hard.bargains

Decision toolPrice board checked Sept. 16, 2026

Keep it, add a card, or buy the whole machine?

Check what context consumes

Name the model and what you run it on. One verdict, the memory behind it and a dated price, including when the answer is to keep what you have.

A card's own memory, or a Mac's total. A Mac leaves a model about two-thirds of 32GB or less and three-quarters above that.

Any budget · Long documentsChange
04 How you will use it

Choose a model and what you run it on.

Four things this can tell you

Keep · Try offloading · Add a card · Buy a machine

The model's weights have to fit in memory, and then your context, your operating system and everything else needs room on top. Those are two different numbers and they decide whether it runs. A third number, memory bandwidth, decides how fast — which is why a 128GB box and a 24GB card can both be right answers.

Every route below leads with the one that costs nothing. With JavaScript switched off, the full board follows: every model, every route that holds it, cheapest first on the total you pay, with anything we found out of stock last, dated prices and the compatibility you have to check.

Jump to a model: Llama 3.1 8B · Qwen3 14B · gpt-oss-20b · Qwen 3.5 27B · Qwen3 30B-A3B · Llama 3.3 70B · gpt-oss-120b · Qwen3-235B-A22B · DeepSeek R1 671B

Fits the weights. What about the conversation?

Model weights are one allocation. Context cache, runtime buffers and other users of the same memory pool need room too. This illustration uses Qwen3-8B's standard full-attention architecture and a 16-bit KV cache, not the model selected in the tool above.

Context alone, one parallel request

4,096 tokens
0.5625 GiB
8,192 tokens
1.125 GiB
16,384 tokens
2.25 GiB
32,768 tokens
4.5 GiB
40,960 tokens
5.625 GiB

Calculated cache payload only: 36 layers × 8 KV heads × 128 dimensions × 2 bytes × keys and values. Padding, runtime allocations and temporary peaks are additional. Weight quantization does not set KV-cache precision.

Complete allocation unknown. Enter weights, runtime allowance and reserve for the selected pool; an omitted number is not zero.

Where the numbers stop being a buying answer

The same GPU memory cannot be counted again as system RAM. For unified memory, reserve space for the OS and applications; dedicated VRAM does not replace the host RAM those need. Do not add separate machines' memory. This example excludes CPU offloading, other architectures, cache quantization, sliding-window attention and changed position scaling. Its context ceiling is the configuration's 40,960 positions, not an extended-context promise.

Even a positive arithmetic margin does not establish runtime support, successful allocation, acceptable latency or model quality. An advertised NPU TOPS figure is not proof that this runtime uses the NPU. Verify the exact runtime, model artifact, drivers, offloading and context on the hardware you will use before treating a price as useful.

Sources checked : Qwen's architecture configuration, Hugging Face's cache-payload method, Ollama context settings and offloading check, Ollama parallel-memory behavior and runtime hardware requirements. The earlier board's weight estimates and dated price receipts remain separately labeled.

Every route, by model

Within a model, every capacity bar is drawn to the same scale. The solid rule is the weights; the dashed rule is the recommended working figure. A bar that stops before the dashed rule will run, and will run out of room. The short meter under each memory figure is memory bandwidth on the same scale — the reason a cheap 128GB box is not a fast one.

Llama 3.1 8B

8BdenseQ4_K_M about 4.6GB

4.5GB holds the weights. 7.5GB is canirun.ai's recommended figure — the difference is your context window, your operating system and everything else you have open.

  1. Keep

    The machine you already own

    $0

    Enough if you already have 4.5GB free for the weights, and comfortable at 7.5GB.

    This is a dense model, so there is no expert-offload trick. It either fits your memory or it runs from disk and crawls.

    The other free routes: run the model you actually need rather than the largest one, or compare a $200-a-month subscription against the hardware in the subscription comparison.

  2. Whole machine

    Mac mini M4, 16GB / 256GB, used

    10.7GBusable of 16GB

    120GB/s

    Holds the 4.5GB of weights and clears the 7.5GB recommended figure with 3.2GB to spare.

    $749median asking price from our crawler, used market, Sept. 2, 2026$625 — lowest single listing, used market, Sept. 2, 2026

    Measured24.11 t/sLlama 2 7B Q4_0

    Also needed: nothing. It is a whole machine.

    Why this number, and what it does not cover
    Measured24.11 t/sLlama 2 7B Q4_0llama-bench tg128, 10-core GPU M4 — llama.cpp Apple Silicon thread

    16GB unified. The cheapest machine here with a measured speed rather than an estimate.

    A complete machine. No CUDA.

  3. One card

    RTX 5060 Ti 16GB, new

    16GB

    448GB/s

    Holds the 4.5GB of weights and clears the 7.5GB recommended figure with 8.5GB to spare.

    $790lowest new listing we quote (Newegg), Newegg, Sept. 15, 2026from $545 — used eBay listings, eBay, Sept. 15, 2026

    Measured90.94 t/sLlama 2 7B Q4_0

    Also needed: a desktop PC with a free slot, 600W supply, 2.5-slot clearance.

    Why this number, and what it does not cover
    Measured90.94 t/sLlama 2 7B Q4_0llama-bench tg128, build 89d1029, flash attention off — llama.cpp CUDA scoreboard
    Measured111.51 t/sgpt-oss-20bllama-bench tg128, build 9ef6b0b, flash attention on — llama.cpp gpt-oss guide

    16GB GDDR7 on a 128-bit interface. Prompt processing fell from 3,839.21 t/s at a 2,048-token prompt to 3,078.06 at 32,768 (llama.cpp gpt-oss guide). Generation speed at depth is not reported.

    One 8-pin cable, or a PCIe Gen 5 cable rated 300W or more. Board length and slot width vary by maker: the exact ASUS Dual board we link is a 2.5-slot card, so measure the clearance under your first slot before ordering it. A different maker's 2-slot 5060 Ti holds the same 16GB.

  4. Whole machine

    Mac mini M6, 16GB / 256GB, new

    10.7GBusable of 16GB

    153GB/s

    Holds the 4.5GB of weights and clears the 7.5GB recommended figure with 3.2GB to spare.

    $899Apple U.S. list price, Apple, Sept. 2, 2026

    Estimateabout 30-31 t/s7B Q4

    Also needed: nothing. It is a whole machine.

    Why this number, and what it does not cover
    Estimateabout 30-31 t/s7B Q4The M4's measured 24.11 t/s scaled by 153/120 GB/s. No M6 has been benchmarked. — llama.cpp Apple Silicon thread

    16GB unified; macOS and your other apps share it. 16GB unified is shared with macOS, so it is tighter than 16GB of dedicated VRAM.

    A complete machine. No CUDA.

  5. One card

    RTX 5070 Ti 16GB, new

    16GB

    896GB/s

    Holds the 4.5GB of weights and clears the 7.5GB recommended figure with 8.5GB to spare.

    $1,150lowest new listing we quote (Newegg), Newegg, Sept. 15, 2026$995 — one new-condition eBay listing, eBay, Sept. 15, 2026

    Measured176.85 t/sLlama 2 7B Q4_0

    Also needed: a desktop PC with a free slot, 750W supply, 2-slot clearance.

    Why this number, and what it does not cover
    Measured176.85 t/sLlama 2 7B Q4_0llama-bench tg128, build 933414c — llama.cpp CUDA scoreboard
    Measured189.45 t/sgpt-oss-20bllama-bench tg128, build a094f38 — llama.cpp gpt-oss guide

    16GB on a 256-bit interface, twice as wide as the 5060 Ti's. Same 16GB as the 5060 Ti. The extra money buys speed, not fit: it generated nearly twice as fast in the 7B test.

    Two 8-pin cables through the included adapter, or a PCIe Gen 5 cable rated 300W or more. Size varies by board.

5 routes here, cheapest first. Change the question.

Qwen3 14B

14BdenseQ4_K_M about 7.7GB

7.8GB holds the weights. 13GB is canirun.ai's recommended figure — the difference is your context window, your operating system and everything else you have open.

  1. Keep

    The machine you already own

    $0

    Enough if you already have 7.8GB free for the weights, and comfortable at 13GB.

    This is a dense model, so there is no expert-offload trick. It either fits your memory or it runs from disk and crawls.

    The other free routes: run the model you actually need rather than the largest one, or compare a $200-a-month subscription against the hardware in the subscription comparison.

  2. Whole machine

    Mac mini M4, 16GB / 256GB, used

    10.7GBusable of 16GB

    120GB/s

    Holds the 7.8GB of weights, but stops 2.3GB short of the 13GB recommended figure. Fine for short prompts; tight once the context fills.

    $749median asking price from our crawler, used market, Sept. 2, 2026$625 — lowest single listing, used market, Sept. 2, 2026

    Measured24.11 t/sLlama 2 7B Q4_0

    Also needed: nothing. It is a whole machine.

    Why this number, and what it does not cover
    Measured24.11 t/sLlama 2 7B Q4_0llama-bench tg128, 10-core GPU M4 — llama.cpp Apple Silicon thread

    16GB unified. The cheapest machine here with a measured speed rather than an estimate.

    A complete machine. No CUDA.

  3. One card

    RTX 5060 Ti 16GB, new

    16GB

    448GB/s

    Holds the 7.8GB of weights and clears the 13GB recommended figure with 3GB to spare.

    $790lowest new listing we quote (Newegg), Newegg, Sept. 15, 2026from $545 — used eBay listings, eBay, Sept. 15, 2026

    Measured90.94 t/sLlama 2 7B Q4_0

    Also needed: a desktop PC with a free slot, 600W supply, 2.5-slot clearance.

    Why this number, and what it does not cover
    Measured90.94 t/sLlama 2 7B Q4_0llama-bench tg128, build 89d1029, flash attention off — llama.cpp CUDA scoreboard
    Measured111.51 t/sgpt-oss-20bllama-bench tg128, build 9ef6b0b, flash attention on — llama.cpp gpt-oss guide

    16GB GDDR7 on a 128-bit interface. Prompt processing fell from 3,839.21 t/s at a 2,048-token prompt to 3,078.06 at 32,768 (llama.cpp gpt-oss guide). Generation speed at depth is not reported.

    One 8-pin cable, or a PCIe Gen 5 cable rated 300W or more. Board length and slot width vary by maker: the exact ASUS Dual board we link is a 2.5-slot card, so measure the clearance under your first slot before ordering it. A different maker's 2-slot 5060 Ti holds the same 16GB.

  4. Whole machine

    Mac mini M6, 16GB / 256GB, new

    10.7GBusable of 16GB

    153GB/s

    Holds the 7.8GB of weights, but stops 2.3GB short of the 13GB recommended figure. Fine for short prompts; tight once the context fills.

    $899Apple U.S. list price, Apple, Sept. 2, 2026

    Estimateabout 30-31 t/s7B Q4

    Also needed: nothing. It is a whole machine.

    Why this number, and what it does not cover
    Estimateabout 30-31 t/s7B Q4The M4's measured 24.11 t/s scaled by 153/120 GB/s. No M6 has been benchmarked. — llama.cpp Apple Silicon thread

    16GB unified; macOS and your other apps share it. 16GB unified is shared with macOS, so it is tighter than 16GB of dedicated VRAM.

    A complete machine. No CUDA.

  5. One card

    RTX 3090 24GB, used

    24GB

    936GB/s

    Holds the 7.8GB of weights and clears the 13GB recommended figure with 11GB to spare.

    $1,032eBay sold-listing price, eBay, Sept. 12, 2026$1,337-$1,463 — fair asking range across 291 active listings, eBay, Sept. 15, 2026

    Measured158.16 t/sLlama 2 7B Q4_0

    Also needed: a desktop PC with a free slot, 750W supply, 3-slot clearance, 313mm of case.

    Check Buy near the sold price, not the ask. The two differ by more than 20% here.

    Why this number, and what it does not cover
    Measured158.16 t/sLlama 2 7B Q4_0llama-bench tg128, build c76b420 — llama.cpp CUDA scoreboard
    Measured161.77 t/sgpt-oss-20bllama-bench tg128, build a094f38 — llama.cpp gpt-oss guide

    24GB on a 384-bit interface. Prompt processing fell from 5,170.56 t/s at a 2,048-token prompt to 3,577.10 at 32,768 (llama.cpp gpt-oss guide). A long agent session pays that cost on every turn.

    Nvidia's Founders Edition uses two 8-pin cables through a 12-pin adapter, takes three slots and is 313mm long. Partner boards differ in all three.

    Buying used: work through the five checks before and after paying.

  6. One card

    RTX 5070 Ti 16GB, new

    16GB

    896GB/s

    Holds the 7.8GB of weights and clears the 13GB recommended figure with 3GB to spare.

    $1,150lowest new listing we quote (Newegg), Newegg, Sept. 15, 2026$995 — one new-condition eBay listing, eBay, Sept. 15, 2026

    Measured176.85 t/sLlama 2 7B Q4_0

    Also needed: a desktop PC with a free slot, 750W supply, 2-slot clearance.

    Why this number, and what it does not cover
    Measured176.85 t/sLlama 2 7B Q4_0llama-bench tg128, build 933414c — llama.cpp CUDA scoreboard
    Measured189.45 t/sgpt-oss-20bllama-bench tg128, build a094f38 — llama.cpp gpt-oss guide

    16GB on a 256-bit interface, twice as wide as the 5060 Ti's. Same 16GB as the 5060 Ti. The extra money buys speed, not fit: it generated nearly twice as fast in the 7B test.

    Two 8-pin cables through the included adapter, or a PCIe Gen 5 cable rated 300W or more. Size varies by board.

  7. Whole machine

    Mac Studio M5 Max, 36GB, new

    27GBusable of 36GB

    460GB/s

    Holds the 7.8GB of weights and clears the 13GB recommended figure with 14GB to spare.

    $2,499Apple U.S. list price, Apple, Aug. 25, 2026

    Estimateabout 21-27 t/s27B dense Q4

    Also needed: nothing. It is a whole machine.

    Why this number, and what it does not cover
    Estimateabout 21-27 t/s27B dense Q4Bandwidth arithmetic: 460GB/s divided by the model's size, times 0.7-0.9. No M5 has been benchmarked. — llama.cpp Apple Silicon thread

    36GB unified. 36GB unified holds a 27B model's 15.5GB weights with real room for context, which a 16GB card does not.

    A complete machine. No CUDA.

  8. One card

    RTX 5090 32GB, new

    32GB

    1,792GB/s

    Holds the 7.8GB of weights and clears the 13GB recommended figure with 19GB to spare.

    $6,700lowest new listing we quote (Newegg), Newegg, Sept. 15, 2026At about $209 a gigabyte, this is the most expensive memory on this page.no eBay offer found — used, eBay, Sept. 15, 2026

    Measured290.02 t/sLlama 2 7B Q4_0

    Also needed: a desktop PC with a free slot, 1,000W supply, 2-slot clearance, 304mm of case.

    Check At $6,700 this is $209 per gigabyte of VRAM, against $43 for a used 3090 at its sold price. Our August memory comparison used a reported street price near $4,700; the $6,700 figure is a Sept. 15 Newegg listing.

    Why this number, and what it does not cover
    Measured290.02 t/sLlama 2 7B Q4_0llama-bench tg128, build 8cf6b42 — llama.cpp CUDA scoreboard
    Measured282.51 t/sgpt-oss-20bllama-bench tg128, build a094f38 — llama.cpp gpt-oss guide

    32GB on a 512-bit interface. The fastest single card here, and still 32GB: a 4-bit 70B model does not fit.

    Four 8-pin cables through an adapter, or one 600W PCIe Gen 5 cable. The Founders Edition takes two slots and is 304mm long.

8 routes here, cheapest first. Change the question.

gpt-oss-20b

21B total, 3.6B activemixture of expertsQ4_K_M about 11.3GB

11.7GB holds the weights. 19.6GB is canirun.ai's recommended figure — the difference is your context window, your operating system and everything else you have open.

  1. Keep

    The machine you already own

    $0

    Enough if you already have 11.7GB free for the weights, and comfortable at 19.6GB.

    This is a mixture-of-experts model, so llama.cpp can keep attention on the GPU and push the expert layers into system RAM. Its gpt-oss guide configures gpt-oss-120b for a card with 8GB using --n-cpu-moe 35. It publishes no speed for that, because the speed is your CPU and your RAM. Two RTX 3090s with 26 expert layers in RAM measured 10-17 t/s. Try it on the machine you own before you spend anything.

    The other free routes: run the model you actually need rather than the largest one, or compare a $200-a-month subscription against the hardware in the subscription comparison.

  2. One card

    RTX 5060 Ti 16GB, new

    16GB

    448GB/s

    Holds the 11.7GB of weights, but stops 3.6GB short of the 19.6GB recommended figure. Fine for short prompts; tight once the context fills.

    $790lowest new listing we quote (Newegg), Newegg, Sept. 15, 2026from $545 — used eBay listings, eBay, Sept. 15, 2026

    Measured111.51 t/sgpt-oss-20b

    Also needed: a desktop PC with a free slot, 600W supply, 2.5-slot clearance.

    Why this number, and what it does not cover
    Measured90.94 t/sLlama 2 7B Q4_0llama-bench tg128, build 89d1029, flash attention off — llama.cpp CUDA scoreboard
    Measured111.51 t/sgpt-oss-20bllama-bench tg128, build 9ef6b0b, flash attention on — llama.cpp gpt-oss guide

    16GB GDDR7 on a 128-bit interface. Prompt processing fell from 3,839.21 t/s at a 2,048-token prompt to 3,078.06 at 32,768 (llama.cpp gpt-oss guide). Generation speed at depth is not reported.

    One 8-pin cable, or a PCIe Gen 5 cable rated 300W or more. Board length and slot width vary by maker: the exact ASUS Dual board we link is a 2.5-slot card, so measure the clearance under your first slot before ordering it. A different maker's 2-slot 5060 Ti holds the same 16GB.

  3. One card

    RTX 3090 24GB, used

    24GB

    936GB/s

    Holds the 11.7GB of weights and clears the 19.6GB recommended figure with 4.4GB to spare.

    $1,032eBay sold-listing price, eBay, Sept. 12, 2026$1,337-$1,463 — fair asking range across 291 active listings, eBay, Sept. 15, 2026

    Measured161.77 t/sgpt-oss-20b

    Also needed: a desktop PC with a free slot, 750W supply, 3-slot clearance, 313mm of case.

    Check Buy near the sold price, not the ask. The two differ by more than 20% here.

    Why this number, and what it does not cover
    Measured158.16 t/sLlama 2 7B Q4_0llama-bench tg128, build c76b420 — llama.cpp CUDA scoreboard
    Measured161.77 t/sgpt-oss-20bllama-bench tg128, build a094f38 — llama.cpp gpt-oss guide

    24GB on a 384-bit interface. Prompt processing fell from 5,170.56 t/s at a 2,048-token prompt to 3,577.10 at 32,768 (llama.cpp gpt-oss guide). A long agent session pays that cost on every turn.

    Nvidia's Founders Edition uses two 8-pin cables through a 12-pin adapter, takes three slots and is 313mm long. Partner boards differ in all three.

    Buying used: work through the five checks before and after paying.

  4. One card

    RTX 5070 Ti 16GB, new

    16GB

    896GB/s

    Holds the 11.7GB of weights, but stops 3.6GB short of the 19.6GB recommended figure. Fine for short prompts; tight once the context fills.

    $1,150lowest new listing we quote (Newegg), Newegg, Sept. 15, 2026$995 — one new-condition eBay listing, eBay, Sept. 15, 2026

    Measured189.45 t/sgpt-oss-20b

    Also needed: a desktop PC with a free slot, 750W supply, 2-slot clearance.

    Why this number, and what it does not cover
    Measured176.85 t/sLlama 2 7B Q4_0llama-bench tg128, build 933414c — llama.cpp CUDA scoreboard
    Measured189.45 t/sgpt-oss-20bllama-bench tg128, build a094f38 — llama.cpp gpt-oss guide

    16GB on a 256-bit interface, twice as wide as the 5060 Ti's. Same 16GB as the 5060 Ti. The extra money buys speed, not fit: it generated nearly twice as fast in the 7B test.

    Two 8-pin cables through the included adapter, or a PCIe Gen 5 cable rated 300W or more. Size varies by board.

  5. Two cards

    Two used RTX 3090s

    48GB

    936GB/s per card, not pooled

    Holds the 11.7GB of weights and clears the 19.6GB recommended figure with 28.4GB to spare.

    $2,064two cards at the Sept. 12 sold price (our arithmetic), eBay, Sept. 12, 2026$2,674-$2,926 — two cards at the Sept. 15 asking range (our arithmetic), eBay, Sept. 15, 2026

    Measured10-17 t/sgpt-oss-120b with 26 expert layers in system RAM

    Also needed: a desktop PC with a free slot, 1,000W supply, 6-slot clearance.

    Why this number, and what it does not cover
    Measured10-17 t/sgpt-oss-120b with 26 expert layers in system RAMQ4_K_M, --n-cpu-moe 26, 64GB DDR4, Ryzen 7 5800X, 38,000-token context, January 2026 — LLM Garage dual-3090 run
    Estimateabout 14-16 t/s70B dense at Q4_K_MCommunity-derived estimate. We found no measured 70B figure for this pair from a source we could check. — Compute Market multi-GPU guide

    24GB + 24GB. Weights split across both cards; this is not one 48GB pool for every runtime. This is the cheapest route to 48GB, and the one most likely to be blocked by your case and power supply rather than your budget.

    700W of rated card power before the rest of the PC. Compute Market recommends at least a 1,000W supply. The board needs two slots spaced far enough apart for both coolers, and the case needs room for both; a Founders Edition alone takes three slots. Check each slot's PCIe lane count in the board manual.

    Buying used: work through the five checks before and after paying.

  6. Whole machine

    Mac Studio M5 Max, 36GB, new

    27GBusable of 36GB

    460GB/s

    Holds the 11.7GB of weights and clears the 19.6GB recommended figure with 7.4GB to spare.

    $2,499Apple U.S. list price, Apple, Aug. 25, 2026

    Estimateabout 21-27 t/s27B dense Q4

    Also needed: nothing. It is a whole machine.

    Why this number, and what it does not cover
    Estimateabout 21-27 t/s27B dense Q4Bandwidth arithmetic: 460GB/s divided by the model's size, times 0.7-0.9. No M5 has been benchmarked. — llama.cpp Apple Silicon thread

    36GB unified. 36GB unified holds a 27B model's 15.5GB weights with real room for context, which a 16GB card does not.

    A complete machine. No CUDA.

  7. Whole machine

    Mac mini M5 Pro, 64GB / 1TB, new

    48GBusable of 64GB

    307GB/s

    Holds the 11.7GB of weights and clears the 19.6GB recommended figure with 28.4GB to spare.

    $3,199Apple U.S. list price, Apple, Aug. 25, 2026

    Estimateabout 5-7 t/s70B dense Q4

    Also needed: nothing. It is a whole machine.

    Check Ships Sept. 22, 2026 per Apple's Aug. 25 quotes. Recheck the date before ordering.

    Why this number, and what it does not cover
    Estimateabout 5-7 t/s70B dense Q4Bandwidth arithmetic: 307GB/s divided by about 40GB, times 0.7-0.9. — llama.cpp Apple Silicon thread

    64GB unified. 64GB clears a 70B model's 39.1GB weights but not its 65.2GB recommended figure. Expect single-digit tokens per second.

    A complete machine. No CUDA.

  8. Whole machine

    Mac Studio M5 Max, 40-core GPU, 64GB, new

    48GBusable of 64GB

    614GB/s

    Holds the 11.7GB of weights and clears the 19.6GB recommended figure with 28.4GB to spare.

    about $3,500rounded configuration price, Apple, Aug. 26, 2026

    Estimateabout 11-14 t/s70B dense Q4

    Also needed: nothing. It is a whole machine.

    Why this number, and what it does not cover
    Estimateabout 11-14 t/s70B dense Q4Bandwidth arithmetic: 614GB/s divided by about 40GB, times 0.7-0.9. — llama.cpp Apple Silicon thread

    64GB unified. Twice the bandwidth of the M5 Pro mini for about $300 more, at the same 64GB.

    A complete machine. No CUDA.

  9. One card

    RTX 5090 32GB, new

    32GB

    1,792GB/s

    Holds the 11.7GB of weights and clears the 19.6GB recommended figure with 12.4GB to spare.

    $6,700lowest new listing we quote (Newegg), Newegg, Sept. 15, 2026At about $209 a gigabyte, this is the most expensive memory on this page.no eBay offer found — used, eBay, Sept. 15, 2026

    Measured282.51 t/sgpt-oss-20b

    Also needed: a desktop PC with a free slot, 1,000W supply, 2-slot clearance, 304mm of case.

    Check At $6,700 this is $209 per gigabyte of VRAM, against $43 for a used 3090 at its sold price. Our August memory comparison used a reported street price near $4,700; the $6,700 figure is a Sept. 15 Newegg listing.

    Why this number, and what it does not cover
    Measured290.02 t/sLlama 2 7B Q4_0llama-bench tg128, build 8cf6b42 — llama.cpp CUDA scoreboard
    Measured282.51 t/sgpt-oss-20bllama-bench tg128, build a094f38 — llama.cpp gpt-oss guide

    32GB on a 512-bit interface. The fastest single card here, and still 32GB: a 4-bit 70B model does not fit.

    Four 8-pin cables through an adapter, or one 600W PCIe Gen 5 cable. The Founders Edition takes two slots and is 304mm long.

9 routes here, cheapest first. Change the question.

Qwen 3.5 27B

27.8BdenseQ4_K_M about 14.7GB

15.5GB holds the weights. 25.9GB is canirun.ai's recommended figure — the difference is your context window, your operating system and everything else you have open.

  1. Keep

    The machine you already own

    $0

    Enough if you already have 15.5GB free for the weights, and comfortable at 25.9GB.

    This is a dense model, so there is no expert-offload trick. It either fits your memory or it runs from disk and crawls.

    The other free routes: run the model you actually need rather than the largest one, or compare a $200-a-month subscription against the hardware in the subscription comparison.

  2. One card

    RTX 5060 Ti 16GB, new

    16GB

    448GB/s

    Holds the 15.5GB of weights, but stops 9.9GB short of the 25.9GB recommended figure. Fine for short prompts; tight once the context fills.

    $790lowest new listing we quote (Newegg), Newegg, Sept. 15, 2026from $545 — used eBay listings, eBay, Sept. 15, 2026

    Measured90.94 t/sLlama 2 7B Q4_0

    Also needed: a desktop PC with a free slot, 600W supply, 2.5-slot clearance.

    Why this number, and what it does not cover
    Measured90.94 t/sLlama 2 7B Q4_0llama-bench tg128, build 89d1029, flash attention off — llama.cpp CUDA scoreboard
    Measured111.51 t/sgpt-oss-20bllama-bench tg128, build 9ef6b0b, flash attention on — llama.cpp gpt-oss guide

    16GB GDDR7 on a 128-bit interface. Prompt processing fell from 3,839.21 t/s at a 2,048-token prompt to 3,078.06 at 32,768 (llama.cpp gpt-oss guide). Generation speed at depth is not reported.

    One 8-pin cable, or a PCIe Gen 5 cable rated 300W or more. Board length and slot width vary by maker: the exact ASUS Dual board we link is a 2.5-slot card, so measure the clearance under your first slot before ordering it. A different maker's 2-slot 5060 Ti holds the same 16GB.

  3. One card

    RTX 3090 24GB, used

    24GB

    936GB/s

    Holds the 15.5GB of weights, but stops 1.9GB short of the 25.9GB recommended figure. Fine for short prompts; tight once the context fills.

    $1,032eBay sold-listing price, eBay, Sept. 12, 2026$1,337-$1,463 — fair asking range across 291 active listings, eBay, Sept. 15, 2026

    Measured158.16 t/sLlama 2 7B Q4_0

    Also needed: a desktop PC with a free slot, 750W supply, 3-slot clearance, 313mm of case.

    Check Buy near the sold price, not the ask. The two differ by more than 20% here.

    Why this number, and what it does not cover
    Measured158.16 t/sLlama 2 7B Q4_0llama-bench tg128, build c76b420 — llama.cpp CUDA scoreboard
    Measured161.77 t/sgpt-oss-20bllama-bench tg128, build a094f38 — llama.cpp gpt-oss guide

    24GB on a 384-bit interface. Prompt processing fell from 5,170.56 t/s at a 2,048-token prompt to 3,577.10 at 32,768 (llama.cpp gpt-oss guide). A long agent session pays that cost on every turn.

    Nvidia's Founders Edition uses two 8-pin cables through a 12-pin adapter, takes three slots and is 313mm long. Partner boards differ in all three.

    Buying used: work through the five checks before and after paying.

  4. One card

    RTX 5070 Ti 16GB, new

    16GB

    896GB/s

    Holds the 15.5GB of weights, but stops 9.9GB short of the 25.9GB recommended figure. Fine for short prompts; tight once the context fills.

    $1,150lowest new listing we quote (Newegg), Newegg, Sept. 15, 2026$995 — one new-condition eBay listing, eBay, Sept. 15, 2026

    Measured176.85 t/sLlama 2 7B Q4_0

    Also needed: a desktop PC with a free slot, 750W supply, 2-slot clearance.

    Why this number, and what it does not cover
    Measured176.85 t/sLlama 2 7B Q4_0llama-bench tg128, build 933414c — llama.cpp CUDA scoreboard
    Measured189.45 t/sgpt-oss-20bllama-bench tg128, build a094f38 — llama.cpp gpt-oss guide

    16GB on a 256-bit interface, twice as wide as the 5060 Ti's. Same 16GB as the 5060 Ti. The extra money buys speed, not fit: it generated nearly twice as fast in the 7B test.

    Two 8-pin cables through the included adapter, or a PCIe Gen 5 cable rated 300W or more. Size varies by board.

  5. Two cards

    Two used RTX 3090s

    48GB

    936GB/s per card, not pooled

    Holds the 15.5GB of weights and clears the 25.9GB recommended figure with 22.1GB to spare.

    $2,064two cards at the Sept. 12 sold price (our arithmetic), eBay, Sept. 12, 2026$2,674-$2,926 — two cards at the Sept. 15 asking range (our arithmetic), eBay, Sept. 15, 2026

    Measured10-17 t/sgpt-oss-120b with 26 expert layers in system RAM

    Also needed: a desktop PC with a free slot, 1,000W supply, 6-slot clearance.

    Why this number, and what it does not cover
    Measured10-17 t/sgpt-oss-120b with 26 expert layers in system RAMQ4_K_M, --n-cpu-moe 26, 64GB DDR4, Ryzen 7 5800X, 38,000-token context, January 2026 — LLM Garage dual-3090 run
    Estimateabout 14-16 t/s70B dense at Q4_K_MCommunity-derived estimate. We found no measured 70B figure for this pair from a source we could check. — Compute Market multi-GPU guide

    24GB + 24GB. Weights split across both cards; this is not one 48GB pool for every runtime. This is the cheapest route to 48GB, and the one most likely to be blocked by your case and power supply rather than your budget.

    700W of rated card power before the rest of the PC. Compute Market recommends at least a 1,000W supply. The board needs two slots spaced far enough apart for both coolers, and the case needs room for both; a Founders Edition alone takes three slots. Check each slot's PCIe lane count in the board manual.

    Buying used: work through the five checks before and after paying.

  6. Whole machine

    Mac Studio M5 Max, 36GB, new

    27GBusable of 36GB

    460GB/s

    Holds the 15.5GB of weights and clears the 25.9GB recommended figure with 1.1GB to spare.

    $2,499Apple U.S. list price, Apple, Aug. 25, 2026

    Estimateabout 21-27 t/s27B dense Q4

    Also needed: nothing. It is a whole machine.

    Why this number, and what it does not cover
    Estimateabout 21-27 t/s27B dense Q4Bandwidth arithmetic: 460GB/s divided by the model's size, times 0.7-0.9. No M5 has been benchmarked. — llama.cpp Apple Silicon thread

    36GB unified. 36GB unified holds a 27B model's 15.5GB weights with real room for context, which a 16GB card does not.

    A complete machine. No CUDA.

  7. Whole machine

    Mac mini M5 Pro, 64GB / 1TB, new

    48GBusable of 64GB

    307GB/s

    Holds the 15.5GB of weights and clears the 25.9GB recommended figure with 22.1GB to spare.

    $3,199Apple U.S. list price, Apple, Aug. 25, 2026

    Estimateabout 5-7 t/s70B dense Q4

    Also needed: nothing. It is a whole machine.

    Check Ships Sept. 22, 2026 per Apple's Aug. 25 quotes. Recheck the date before ordering.

    Why this number, and what it does not cover
    Estimateabout 5-7 t/s70B dense Q4Bandwidth arithmetic: 307GB/s divided by about 40GB, times 0.7-0.9. — llama.cpp Apple Silicon thread

    64GB unified. 64GB clears a 70B model's 39.1GB weights but not its 65.2GB recommended figure. Expect single-digit tokens per second.

    A complete machine. No CUDA.

  8. Whole machine

    Mac Studio M5 Max, 40-core GPU, 64GB, new

    48GBusable of 64GB

    614GB/s

    Holds the 15.5GB of weights and clears the 25.9GB recommended figure with 22.1GB to spare.

    about $3,500rounded configuration price, Apple, Aug. 26, 2026

    Estimateabout 11-14 t/s70B dense Q4

    Also needed: nothing. It is a whole machine.

    Why this number, and what it does not cover
    Estimateabout 11-14 t/s70B dense Q4Bandwidth arithmetic: 614GB/s divided by about 40GB, times 0.7-0.9. — llama.cpp Apple Silicon thread

    64GB unified. Twice the bandwidth of the M5 Pro mini for about $300 more, at the same 64GB.

    A complete machine. No CUDA.

  9. Whole machine

    Mac Studio M5 Ultra, 96GB, new

    72GBusable of 96GB

    1.2TB/s

    Holds the 15.5GB of weights and clears the 25.9GB recommended figure with 46.1GB to spare.

    $5,499Apple U.S. list price, Apple, Aug. 25, 2026

    Estimateabout 104 t/sgpt-oss-120b

    Also needed: nothing. It is a whole machine.

    Why this number, and what it does not cover
    Estimateabout 104 t/sgpt-oss-120bThe M3 Ultra's served 69.39 t/s scaled by 1.5. No M5 Ultra has been benchmarked. — llama.cpp Apple Silicon thread

    96GB unified. 96GB clears gpt-oss-120b's 65.4GB weights but not its 109GB recommended figure.

    A complete machine. No CUDA.

  10. One card

    RTX 5090 32GB, new

    32GB

    1,792GB/s

    Holds the 15.5GB of weights and clears the 25.9GB recommended figure with 6.1GB to spare.

    $6,700lowest new listing we quote (Newegg), Newegg, Sept. 15, 2026At about $209 a gigabyte, this is the most expensive memory on this page.no eBay offer found — used, eBay, Sept. 15, 2026

    Measured290.02 t/sLlama 2 7B Q4_0

    Also needed: a desktop PC with a free slot, 1,000W supply, 2-slot clearance, 304mm of case.

    Check At $6,700 this is $209 per gigabyte of VRAM, against $43 for a used 3090 at its sold price. Our August memory comparison used a reported street price near $4,700; the $6,700 figure is a Sept. 15 Newegg listing.

    Why this number, and what it does not cover
    Measured290.02 t/sLlama 2 7B Q4_0llama-bench tg128, build 8cf6b42 — llama.cpp CUDA scoreboard
    Measured282.51 t/sgpt-oss-20bllama-bench tg128, build a094f38 — llama.cpp gpt-oss guide

    32GB on a 512-bit interface. The fastest single card here, and still 32GB: a 4-bit 70B model does not fit.

    Four 8-pin cables through an adapter, or one 600W PCIe Gen 5 cable. The Founders Edition takes two slots and is 304mm long.

10 routes here, cheapest first. Change the question.

Qwen3 30B-A3B

30B total, 3.3B activemixture of expertsQ4_K_M about 15.9GB

16.8GB holds the weights. 27.9GB is canirun.ai's recommended figure — the difference is your context window, your operating system and everything else you have open.

  1. Keep

    The machine you already own

    $0

    Enough if you already have 16.8GB free for the weights, and comfortable at 27.9GB.

    This is a mixture-of-experts model, so llama.cpp can keep attention on the GPU and push the expert layers into system RAM. Its gpt-oss guide configures gpt-oss-120b for a card with 8GB using --n-cpu-moe 35. It publishes no speed for that, because the speed is your CPU and your RAM. Two RTX 3090s with 26 expert layers in RAM measured 10-17 t/s. Try it on the machine you own before you spend anything.

    The other free routes: run the model you actually need rather than the largest one, or compare a $200-a-month subscription against the hardware in the subscription comparison.

  2. One card

    RTX 3090 24GB, used

    24GB

    936GB/s

    Holds the 16.8GB of weights, but stops 3.9GB short of the 27.9GB recommended figure. Fine for short prompts; tight once the context fills.

    $1,032eBay sold-listing price, eBay, Sept. 12, 2026$1,337-$1,463 — fair asking range across 291 active listings, eBay, Sept. 15, 2026

    Measured158.16 t/sLlama 2 7B Q4_0

    Also needed: a desktop PC with a free slot, 750W supply, 3-slot clearance, 313mm of case.

    Check Buy near the sold price, not the ask. The two differ by more than 20% here.

    Why this number, and what it does not cover
    Measured158.16 t/sLlama 2 7B Q4_0llama-bench tg128, build c76b420 — llama.cpp CUDA scoreboard
    Measured161.77 t/sgpt-oss-20bllama-bench tg128, build a094f38 — llama.cpp gpt-oss guide

    24GB on a 384-bit interface. Prompt processing fell from 5,170.56 t/s at a 2,048-token prompt to 3,577.10 at 32,768 (llama.cpp gpt-oss guide). A long agent session pays that cost on every turn.

    Nvidia's Founders Edition uses two 8-pin cables through a 12-pin adapter, takes three slots and is 313mm long. Partner boards differ in all three.

    Buying used: work through the five checks before and after paying.

  3. Two cards

    Two used RTX 3090s

    48GB

    936GB/s per card, not pooled

    Holds the 16.8GB of weights and clears the 27.9GB recommended figure with 20.1GB to spare.

    $2,064two cards at the Sept. 12 sold price (our arithmetic), eBay, Sept. 12, 2026$2,674-$2,926 — two cards at the Sept. 15 asking range (our arithmetic), eBay, Sept. 15, 2026

    Measured10-17 t/sgpt-oss-120b with 26 expert layers in system RAM

    Also needed: a desktop PC with a free slot, 1,000W supply, 6-slot clearance.

    Why this number, and what it does not cover
    Measured10-17 t/sgpt-oss-120b with 26 expert layers in system RAMQ4_K_M, --n-cpu-moe 26, 64GB DDR4, Ryzen 7 5800X, 38,000-token context, January 2026 — LLM Garage dual-3090 run
    Estimateabout 14-16 t/s70B dense at Q4_K_MCommunity-derived estimate. We found no measured 70B figure for this pair from a source we could check. — Compute Market multi-GPU guide

    24GB + 24GB. Weights split across both cards; this is not one 48GB pool for every runtime. This is the cheapest route to 48GB, and the one most likely to be blocked by your case and power supply rather than your budget.

    700W of rated card power before the rest of the PC. Compute Market recommends at least a 1,000W supply. The board needs two slots spaced far enough apart for both coolers, and the case needs room for both; a Founders Edition alone takes three slots. Check each slot's PCIe lane count in the board manual.

    Buying used: work through the five checks before and after paying.

  4. Whole machine

    Mac Studio M5 Max, 36GB, new

    27GBusable of 36GB

    460GB/s

    Holds the 16.8GB of weights, but stops 0.9GB short of the 27.9GB recommended figure. Fine for short prompts; tight once the context fills.

    $2,499Apple U.S. list price, Apple, Aug. 25, 2026

    Estimateabout 21-27 t/s27B dense Q4

    Also needed: nothing. It is a whole machine.

    Why this number, and what it does not cover
    Estimateabout 21-27 t/s27B dense Q4Bandwidth arithmetic: 460GB/s divided by the model's size, times 0.7-0.9. No M5 has been benchmarked. — llama.cpp Apple Silicon thread

    36GB unified. 36GB unified holds a 27B model's 15.5GB weights with real room for context, which a 16GB card does not.

    A complete machine. No CUDA.

  5. Whole machine

    Mac mini M5 Pro, 64GB / 1TB, new

    48GBusable of 64GB

    307GB/s

    Holds the 16.8GB of weights and clears the 27.9GB recommended figure with 20.1GB to spare.

    $3,199Apple U.S. list price, Apple, Aug. 25, 2026

    Estimateabout 5-7 t/s70B dense Q4

    Also needed: nothing. It is a whole machine.

    Check Ships Sept. 22, 2026 per Apple's Aug. 25 quotes. Recheck the date before ordering.

    Why this number, and what it does not cover
    Estimateabout 5-7 t/s70B dense Q4Bandwidth arithmetic: 307GB/s divided by about 40GB, times 0.7-0.9. — llama.cpp Apple Silicon thread

    64GB unified. 64GB clears a 70B model's 39.1GB weights but not its 65.2GB recommended figure. Expect single-digit tokens per second.

    A complete machine. No CUDA.

  6. Whole machine

    Mac Studio M5 Max, 40-core GPU, 64GB, new

    48GBusable of 64GB

    614GB/s

    Holds the 16.8GB of weights and clears the 27.9GB recommended figure with 20.1GB to spare.

    about $3,500rounded configuration price, Apple, Aug. 26, 2026

    Estimateabout 11-14 t/s70B dense Q4

    Also needed: nothing. It is a whole machine.

    Why this number, and what it does not cover
    Estimateabout 11-14 t/s70B dense Q4Bandwidth arithmetic: 614GB/s divided by about 40GB, times 0.7-0.9. — llama.cpp Apple Silicon thread

    64GB unified. Twice the bandwidth of the M5 Pro mini for about $300 more, at the same 64GB.

    A complete machine. No CUDA.

  7. Whole machine

    Mac Studio M5 Ultra, 96GB, new

    72GBusable of 96GB

    1.2TB/s

    Holds the 16.8GB of weights and clears the 27.9GB recommended figure with 44.1GB to spare.

    $5,499Apple U.S. list price, Apple, Aug. 25, 2026

    Estimateabout 104 t/sgpt-oss-120b

    Also needed: nothing. It is a whole machine.

    Why this number, and what it does not cover
    Estimateabout 104 t/sgpt-oss-120bThe M3 Ultra's served 69.39 t/s scaled by 1.5. No M5 Ultra has been benchmarked. — llama.cpp Apple Silicon thread

    96GB unified. 96GB clears gpt-oss-120b's 65.4GB weights but not its 109GB recommended figure.

    A complete machine. No CUDA.

  8. One card

    RTX 5090 32GB, new

    32GB

    1,792GB/s

    Holds the 16.8GB of weights and clears the 27.9GB recommended figure with 4.1GB to spare.

    $6,700lowest new listing we quote (Newegg), Newegg, Sept. 15, 2026At about $209 a gigabyte, this is the most expensive memory on this page.no eBay offer found — used, eBay, Sept. 15, 2026

    Measured290.02 t/sLlama 2 7B Q4_0

    Also needed: a desktop PC with a free slot, 1,000W supply, 2-slot clearance, 304mm of case.

    Check At $6,700 this is $209 per gigabyte of VRAM, against $43 for a used 3090 at its sold price. Our August memory comparison used a reported street price near $4,700; the $6,700 figure is a Sept. 15 Newegg listing.

    Why this number, and what it does not cover
    Measured290.02 t/sLlama 2 7B Q4_0llama-bench tg128, build 8cf6b42 — llama.cpp CUDA scoreboard
    Measured282.51 t/sgpt-oss-20bllama-bench tg128, build a094f38 — llama.cpp gpt-oss guide

    32GB on a 512-bit interface. The fastest single card here, and still 32GB: a 4-bit 70B model does not fit.

    Four 8-pin cables through an adapter, or one 600W PCIe Gen 5 cable. The Founders Edition takes two slots and is 304mm long.

8 routes here, cheapest first. Change the question.

Llama 3.3 70B

70BdenseQ4_K_M about 36.4GB

39.1GB holds the weights. 65.2GB is canirun.ai's recommended figure — the difference is your context window, your operating system and everything else you have open.

  1. Keep

    The machine you already own

    $0

    Enough if you already have 39.1GB free for the weights, and comfortable at 65.2GB.

    This is a dense model, so there is no expert-offload trick. It either fits your memory or it runs from disk and crawls.

    The other free routes: run the model you actually need rather than the largest one, or compare a $200-a-month subscription against the hardware in the subscription comparison.

  2. Two cards

    Two used RTX 3090s

    48GB

    936GB/s per card, not pooled

    Holds the 39.1GB of weights, but stops 17.2GB short of the 65.2GB recommended figure. Fine for short prompts; tight once the context fills.

    $2,064two cards at the Sept. 12 sold price (our arithmetic), eBay, Sept. 12, 2026$2,674-$2,926 — two cards at the Sept. 15 asking range (our arithmetic), eBay, Sept. 15, 2026

    Estimateabout 14-16 t/s70B dense at Q4_K_M

    Also needed: a desktop PC with a free slot, 1,000W supply, 6-slot clearance.

    Why this number, and what it does not cover
    Measured10-17 t/sgpt-oss-120b with 26 expert layers in system RAMQ4_K_M, --n-cpu-moe 26, 64GB DDR4, Ryzen 7 5800X, 38,000-token context, January 2026 — LLM Garage dual-3090 run
    Estimateabout 14-16 t/s70B dense at Q4_K_MCommunity-derived estimate. We found no measured 70B figure for this pair from a source we could check. — Compute Market multi-GPU guide

    24GB + 24GB. Weights split across both cards; this is not one 48GB pool for every runtime. This is the cheapest route to 48GB, and the one most likely to be blocked by your case and power supply rather than your budget.

    700W of rated card power before the rest of the PC. Compute Market recommends at least a 1,000W supply. The board needs two slots spaced far enough apart for both coolers, and the case needs room for both; a Founders Edition alone takes three slots. Check each slot's PCIe lane count in the board manual.

    Buying used: work through the five checks before and after paying.

  3. Whole machine

    Mac mini M5 Pro, 64GB / 1TB, new

    48GBusable of 64GB

    307GB/s

    Holds the 39.1GB of weights, but stops 17.2GB short of the 65.2GB recommended figure. Fine for short prompts; tight once the context fills.

    $3,199Apple U.S. list price, Apple, Aug. 25, 2026

    Estimateabout 5-7 t/s70B dense Q4

    Also needed: nothing. It is a whole machine.

    Check Ships Sept. 22, 2026 per Apple's Aug. 25 quotes. Recheck the date before ordering.

    Why this number, and what it does not cover
    Estimateabout 5-7 t/s70B dense Q4Bandwidth arithmetic: 307GB/s divided by about 40GB, times 0.7-0.9. — llama.cpp Apple Silicon thread

    64GB unified. 64GB clears a 70B model's 39.1GB weights but not its 65.2GB recommended figure. Expect single-digit tokens per second.

    A complete machine. No CUDA.

  4. Whole machine

    GMKtec EVO-X2, Ryzen AI Max+ 395, 128GB

    128GB

    256GB/s

    Holds the 39.1GB of weights and clears the 65.2GB recommended figure with 62.8GB to spare.

    $3,499maker's own store price, 128GB + 1TB, GMKtec, Sept. 16, 2026Both 128GB configurations were listed available on GMKtec's own store on Sept. 16, 2026. The extra 1TB of SSD costs $150 and buys no extra memory for a model.$3,650 — 128GB + 2TB, the maker's price for the larger drive, GMKtec, Sept. 16, 2026

    Unknownno speed we can cite

    Also needed: nothing. It is a whole machine.

    Stock both 128GB configurations available (GMKtec's own store, Sept. 16, 2026)

    Check The 64GB version of this machine looks almost identical in a listing and will not hold a 120B model. Confirm 128GB before paying.

    Why this number, and what it does not cover
    Unknownno measured speedWe have no measured local-model speed for this machine and are not estimating one from bandwidth alone.

    128GB LPDDR5X shared between CPU and integrated GPU. 128GB holds gpt-oss-120b's 65.4GB minimum outright, with room for its 109GB recommended figure. Capacity is the reason to buy it; speed is unproven here. It arrives complete: memory, a 1TB or 2TB drive and a case, with nothing else to add.

    A complete machine. No power supply, case or slot to check. No CUDA: llama.cpp and Ollama run on ROCm or Vulkan, and CUDA-only tools do not.

  5. Whole machine

    Mac Studio M5 Max, 40-core GPU, 64GB, new

    48GBusable of 64GB

    614GB/s

    Holds the 39.1GB of weights, but stops 17.2GB short of the 65.2GB recommended figure. Fine for short prompts; tight once the context fills.

    about $3,500rounded configuration price, Apple, Aug. 26, 2026

    Estimateabout 11-14 t/s70B dense Q4

    Also needed: nothing. It is a whole machine.

    Why this number, and what it does not cover
    Estimateabout 11-14 t/s70B dense Q4Bandwidth arithmetic: 614GB/s divided by about 40GB, times 0.7-0.9. — llama.cpp Apple Silicon thread

    64GB unified. Twice the bandwidth of the M5 Pro mini for about $300 more, at the same 64GB.

    A complete machine. No CUDA.

  6. Whole machine

    Nvidia DGX Spark, 128GB

    128GB

    273GB/s

    Holds the 39.1GB of weights and clears the 65.2GB recommended figure with 62.8GB to spare.

    $4,699reported price, not rechecked, reported in our August comparison, Aug. 26, 2026 stale

    Unknownno speed we can cite

    Also needed: nothing. It is a whole machine.

    Check The $4,699 figure was reported in August and has not been rechecked since. Treat it as stale.

    Why this number, and what it does not cover
    Unknownno measured speedNo measured local-model speed cited here.

    128GB unified. The CUDA option at 128GB. Its 273GB/s is about a third of a used M1 Ultra's bandwidth.

    A complete machine, and the only 128GB box here that runs CUDA.

  7. Whole machine

    Mac Studio M1 Ultra, 128GB, used

    96GBusable of 128GB

    800GB/s

    Holds the 39.1GB of weights and clears the 65.2GB recommended figure with 30.8GB to spare.

    $5,249median asking price from our crawler, used market, Aug. 25, 2026

    Measured79.68 t/sgpt-oss-120b

    Also needed: nothing. It is a whole machine.

    Why this number, and what it does not cover
    Measured79.68 t/sgpt-oss-120bMeasured on an M2 Ultra 192GB, one generation newer than this machine but at the same 800GB/s class bandwidth. — llama.cpp gpt-oss guide

    128GB unified. Three times the memory bandwidth of a 128GB Strix Halo box, at about $41 per gigabyte.

    A complete machine. No CUDA.

  8. Whole machine

    Mac Studio M5 Ultra, 96GB, new

    72GBusable of 96GB

    1.2TB/s

    Holds the 39.1GB of weights and clears the 65.2GB recommended figure with 6.8GB to spare.

    $5,499Apple U.S. list price, Apple, Aug. 25, 2026

    Estimateabout 104 t/sgpt-oss-120b

    Also needed: nothing. It is a whole machine.

    Why this number, and what it does not cover
    Estimateabout 104 t/sgpt-oss-120bThe M3 Ultra's served 69.39 t/s scaled by 1.5. No M5 Ultra has been benchmarked. — llama.cpp Apple Silicon thread

    96GB unified. 96GB clears gpt-oss-120b's 65.4GB weights but not its 109GB recommended figure.

    A complete machine. No CUDA.

  9. Whole machine

    Mac Studio M3 Ultra, 256GB, used

    192GBusable of 256GB

    819GB/s

    Holds the 39.1GB of weights and clears the 65.2GB recommended figure with 126.8GB to spare.

    $9,999median asking price from our crawler, used market, Aug. 25, 2026

    Measured16 t/s as GGUF, 24 under MLXQwen3-235B-A22B

    Also needed: nothing. It is a whole machine.

    Why this number, and what it does not cover
    Measured16 t/s as GGUF, 24 under MLXQwen3-235B-A22BMeasured in LM Studio on a 512GB M3 Ultra — MacStories early Mac Studio AI benchmarks

    256GB unified. 256GB clears Qwen3-235B's 131.3GB weights; its 218.9GB recommended figure is close to the ceiling.

    A complete machine. No CUDA.

  10. Whole machine

    Framework Desktop, 128GB

    128GB

    256GB/s

    Holds the 39.1GB of weights and clears the 65.2GB recommended figure with 62.8GB to spare.

    $3,449maker's own store price, system only, Framework, Sept. 16, 2026$3,449 is the DIY Edition system without a drive. Framework's cheapest 500GB option adds $135, so a working machine is about $3,584 before an operating system.

    Unknownno speed we can cite

    Also needed: a drive — the cheapest 500GB option on Framework's own configurator adds $135.

    Stock out of stock (Framework's store, rechecked Sept. 16, 2026)

    Check All three Framework Desktop systems, including this 128GB one, were listed out of stock when we reread the configurator on Sept. 16, 2026.

    Why this number, and what it does not cover
    Unknownno measured speedNo measured local-model speed cited here.

    128GB shared, soldered. Same 128GB and same 256GB/s as the EVO-X2. It is $50 cheaper on paper and then needs a drive, which the EVO-X2 includes — and on Sept. 16 you could not buy it.

    A complete machine apart from the drive you add yourself. No CUDA.

    • Framework Desktop, 128GBFramework · exact product · where the price above was readOpens with no system chosen: choose Max+ 395 - 128GBThis is the price and the out-of-stock state we read ourselves.

10 routes here, cheapest first. Change the question.

gpt-oss-120b

117B total, 5.1B activemixture of expertsQ4_K_M about 60.4GB

65.4GB holds the weights. 109GB is canirun.ai's recommended figure — the difference is your context window, your operating system and everything else you have open.

  1. Keep

    The machine you already own

    $0

    Enough if you already have 65.4GB free for the weights, and comfortable at 109GB.

    This is a mixture-of-experts model, so llama.cpp can keep attention on the GPU and push the expert layers into system RAM. Its gpt-oss guide configures gpt-oss-120b for a card with 8GB using --n-cpu-moe 35. It publishes no speed for that, because the speed is your CPU and your RAM. Two RTX 3090s with 26 expert layers in RAM measured 10-17 t/s. Try it on the machine you own before you spend anything.

    The other free routes: run the model you actually need rather than the largest one, or compare a $200-a-month subscription against the hardware in the subscription comparison.

  2. Whole machine

    GMKtec EVO-X2, Ryzen AI Max+ 395, 128GB

    128GB

    256GB/s

    Holds the 65.4GB of weights and clears the 109GB recommended figure with 19GB to spare.

    $3,499maker's own store price, 128GB + 1TB, GMKtec, Sept. 16, 2026Both 128GB configurations were listed available on GMKtec's own store on Sept. 16, 2026. The extra 1TB of SSD costs $150 and buys no extra memory for a model.$3,650 — 128GB + 2TB, the maker's price for the larger drive, GMKtec, Sept. 16, 2026

    Unknownno speed we can cite

    Also needed: nothing. It is a whole machine.

    Stock both 128GB configurations available (GMKtec's own store, Sept. 16, 2026)

    Check The 64GB version of this machine looks almost identical in a listing and will not hold a 120B model. Confirm 128GB before paying.

    Why this number, and what it does not cover
    Unknownno measured speedWe have no measured local-model speed for this machine and are not estimating one from bandwidth alone.

    128GB LPDDR5X shared between CPU and integrated GPU. 128GB holds gpt-oss-120b's 65.4GB minimum outright, with room for its 109GB recommended figure. Capacity is the reason to buy it; speed is unproven here. It arrives complete: memory, a 1TB or 2TB drive and a case, with nothing else to add.

    A complete machine. No power supply, case or slot to check. No CUDA: llama.cpp and Ollama run on ROCm or Vulkan, and CUDA-only tools do not.

  3. Whole machine

    Nvidia DGX Spark, 128GB

    128GB

    273GB/s

    Holds the 65.4GB of weights and clears the 109GB recommended figure with 19GB to spare.

    $4,699reported price, not rechecked, reported in our August comparison, Aug. 26, 2026 stale

    Unknownno speed we can cite

    Also needed: nothing. It is a whole machine.

    Check The $4,699 figure was reported in August and has not been rechecked since. Treat it as stale.

    Why this number, and what it does not cover
    Unknownno measured speedNo measured local-model speed cited here.

    128GB unified. The CUDA option at 128GB. Its 273GB/s is about a third of a used M1 Ultra's bandwidth.

    A complete machine, and the only 128GB box here that runs CUDA.

  4. Whole machine

    Mac Studio M1 Ultra, 128GB, used

    96GBusable of 128GB

    800GB/s

    Holds the 65.4GB of weights, but stops 13GB short of the 109GB recommended figure. Fine for short prompts; tight once the context fills.

    $5,249median asking price from our crawler, used market, Aug. 25, 2026

    Measured79.68 t/sgpt-oss-120b

    Also needed: nothing. It is a whole machine.

    Why this number, and what it does not cover
    Measured79.68 t/sgpt-oss-120bMeasured on an M2 Ultra 192GB, one generation newer than this machine but at the same 800GB/s class bandwidth. — llama.cpp gpt-oss guide

    128GB unified. Three times the memory bandwidth of a 128GB Strix Halo box, at about $41 per gigabyte.

    A complete machine. No CUDA.

  5. Whole machine

    Mac Studio M5 Ultra, 96GB, new

    72GBusable of 96GB

    1.2TB/s

    Holds the 65.4GB of weights, but stops 37GB short of the 109GB recommended figure. Fine for short prompts; tight once the context fills.

    $5,499Apple U.S. list price, Apple, Aug. 25, 2026

    Estimateabout 104 t/sgpt-oss-120b

    Also needed: nothing. It is a whole machine.

    Why this number, and what it does not cover
    Estimateabout 104 t/sgpt-oss-120bThe M3 Ultra's served 69.39 t/s scaled by 1.5. No M5 Ultra has been benchmarked. — llama.cpp Apple Silicon thread

    96GB unified. 96GB clears gpt-oss-120b's 65.4GB weights but not its 109GB recommended figure.

    A complete machine. No CUDA.

  6. Whole machine

    Mac Studio M3 Ultra, 256GB, used

    192GBusable of 256GB

    819GB/s

    Holds the 65.4GB of weights and clears the 109GB recommended figure with 83GB to spare.

    $9,999median asking price from our crawler, used market, Aug. 25, 2026

    Measured16 t/s as GGUF, 24 under MLXQwen3-235B-A22B

    Also needed: nothing. It is a whole machine.

    Why this number, and what it does not cover
    Measured16 t/s as GGUF, 24 under MLXQwen3-235B-A22BMeasured in LM Studio on a 512GB M3 Ultra — MacStories early Mac Studio AI benchmarks

    256GB unified. 256GB clears Qwen3-235B's 131.3GB weights; its 218.9GB recommended figure is close to the ceiling.

    A complete machine. No CUDA.

  7. Whole machine

    Framework Desktop, 128GB

    128GB

    256GB/s

    Holds the 65.4GB of weights and clears the 109GB recommended figure with 19GB to spare.

    $3,449maker's own store price, system only, Framework, Sept. 16, 2026$3,449 is the DIY Edition system without a drive. Framework's cheapest 500GB option adds $135, so a working machine is about $3,584 before an operating system.

    Unknownno speed we can cite

    Also needed: a drive — the cheapest 500GB option on Framework's own configurator adds $135.

    Stock out of stock (Framework's store, rechecked Sept. 16, 2026)

    Check All three Framework Desktop systems, including this 128GB one, were listed out of stock when we reread the configurator on Sept. 16, 2026.

    Why this number, and what it does not cover
    Unknownno measured speedNo measured local-model speed cited here.

    128GB shared, soldered. Same 128GB and same 256GB/s as the EVO-X2. It is $50 cheaper on paper and then needs a drive, which the EVO-X2 includes — and on Sept. 16 you could not buy it.

    A complete machine apart from the drive you add yourself. No CUDA.

    • Framework Desktop, 128GBFramework · exact product · where the price above was readOpens with no system chosen: choose Max+ 395 - 128GBThis is the price and the out-of-stock state we read ourselves.

7 routes here, cheapest first. Change the question.

Qwen3-235B-A22B

235B total, 22B activemixture of expertsQ4_K_M about 120.9GB

131.3GB holds the weights. 218.9GB is canirun.ai's recommended figure — the difference is your context window, your operating system and everything else you have open.

  1. Keep

    The machine you already own

    $0

    Enough if you already have 131.3GB free for the weights, and comfortable at 218.9GB.

    This is a mixture-of-experts model, so llama.cpp can keep attention on the GPU and push the expert layers into system RAM. Its gpt-oss guide configures gpt-oss-120b for a card with 8GB using --n-cpu-moe 35. It publishes no speed for that, because the speed is your CPU and your RAM. Two RTX 3090s with 26 expert layers in RAM measured 10-17 t/s. Try it on the machine you own before you spend anything.

    The other free routes: run the model you actually need rather than the largest one, or compare a $200-a-month subscription against the hardware in the subscription comparison.

  2. Whole machine

    Mac Studio M3 Ultra, 256GB, used

    192GBusable of 256GB

    819GB/s

    Holds the 131.3GB of weights, but stops 26.9GB short of the 218.9GB recommended figure. Fine for short prompts; tight once the context fills.

    $9,999median asking price from our crawler, used market, Aug. 25, 2026

    Measured16 t/s as GGUF, 24 under MLXQwen3-235B-A22B

    Also needed: nothing. It is a whole machine.

    Why this number, and what it does not cover
    Measured16 t/s as GGUF, 24 under MLXQwen3-235B-A22BMeasured in LM Studio on a 512GB M3 Ultra — MacStories early Mac Studio AI benchmarks

    256GB unified. 256GB clears Qwen3-235B's 131.3GB weights; its 218.9GB recommended figure is close to the ceiling.

    A complete machine. No CUDA.

  3. Whole machine

    Mac Studio M3 Ultra, 512GB, used

    384GBusable of 512GB

    819GB/s

    Holds the 131.3GB of weights and clears the 218.9GB recommended figure with 165.1GB to spare.

    $17,448median asking price from our crawler, used market, Aug. 25, 2026

    Measured18.11 t/s under MLX, 16.08 in a second runDeepSeek R1 671B at 4-bit

    Also needed: nothing. It is a whole machine.

    Why this number, and what it does not cover
    Measured18.11 t/s under MLX, 16.08 in a second runDeepSeek R1 671B at 4-bitMeasured on a 512GB M3 Ultra. A separate run with an 11,000-token prompt returned 4-11 t/s. — Hostbor M3 Ultra test

    512GB unified. 512GB clears DeepSeek R1's 375GB weights. Its 624.9GB recommended figure exceeds every single machine here.

    A complete machine. No CUDA.

3 routes here, cheapest first. Change the question.

DeepSeek R1 671B

671B total, 37B activemixture of expertsQ4_K_M about 344.2GB

375GB holds the weights. 624.9GB is canirun.ai's recommended figure — the difference is your context window, your operating system and everything else you have open.

  1. Keep

    The machine you already own

    $0

    Enough if you already have 375GB free for the weights, and comfortable at 624.9GB.

    This is a mixture-of-experts model, so llama.cpp can keep attention on the GPU and push the expert layers into system RAM. Its gpt-oss guide configures gpt-oss-120b for a card with 8GB using --n-cpu-moe 35. It publishes no speed for that, because the speed is your CPU and your RAM. Two RTX 3090s with 26 expert layers in RAM measured 10-17 t/s. Try it on the machine you own before you spend anything.

    The other free routes: run the model you actually need rather than the largest one, or compare a $200-a-month subscription against the hardware in the subscription comparison.

  2. Whole machine

    Mac Studio M3 Ultra, 512GB, used

    384GBusable of 512GB

    819GB/s

    Holds the 375GB of weights, but stops 240.9GB short of the 624.9GB recommended figure. Fine for short prompts; tight once the context fills.

    $17,448median asking price from our crawler, used market, Aug. 25, 2026

    Measured18.11 t/s under MLX, 16.08 in a second runDeepSeek R1 671B at 4-bit

    Also needed: nothing. It is a whole machine.

    Why this number, and what it does not cover
    Measured18.11 t/s under MLX, 16.08 in a second runDeepSeek R1 671B at 4-bitMeasured on a 512GB M3 Ultra. A separate run with an 11,000-token prompt returned 4-11 t/s. — Hostbor M3 Ultra test

    512GB unified. 512GB clears DeepSeek R1's 375GB weights. Its 624.9GB recommended figure exceeds every single machine here.

    A complete machine. No CUDA.

2 routes here, cheapest first. Change the question.

Five checks before and after paying for a used card
  1. Match the exact board, not the chip name. Length, slot width and power connectors differ by maker; Nvidia's 3090 Founders Edition is 313mm long, takes three slots and uses two 8-pin cables through a 12-pin adapter.
  2. Ask why it is being sold, and for a photo or video of the card running with the date visible.
  3. Read the return terms before paying. Without returns, a remote purchase leaves any defect with you.
  4. Test it the day it arrives: loop a demanding benchmark while watching temperatures, fan speed and power, then load the model you bought it for at the context length you use.
  5. Do not read a passing test as a history. Mining runs cards at full power for long stretches; a clean benchmark shows the card works now, not how it was treated.

Checks 2, 4 and 5 follow Tom's Hardware's used-GPU guidance.

Sources and assumptions

Memory figures are canirun.ai's estimates at roughly 4-bit, each model page opened on Sept. 15, 2026. They are not our measurements. Memory bandwidth is a maker or tracker specification, not a speed we measured. A speed is labeled measured only when it names a benchmark run, its build and its settings; everything else is labeled an estimate and shows its arithmetic.

The 2 Amazon product configurations
  • RTX 5060 Ti 16GB, new — ASIN B0F7WB6LSH. Opened Sept. 15, 2026 and reopened Sept. 16: a live listing titled ASUS Dual GeForce RTX 5060 Ti 16GB GDDR7 OC Edition. Its own description calls it a 2.5-slot design. One specific board, not the listing the dated price above came from. No price or stock is shown here for it; check both on the listing.
  • GMKtec EVO-X2, Ryzen AI Max+ 395, 128GB — ASIN B0F53MLYQ6. Opened Sept. 15, 2026 and reopened Sept. 16: a live listing titled GMKtec EVO-X2 AI Mini PC Ryzen AI Max+ 395 128GB LPDDR5X 2TB SSD. The 2TB configuration, not the 1TB priced above: the difference is storage, not memory for a model. No price or stock is shown here for this listing.

An ASIN resolving is not evidence of stock or a fair price. Check both on the listing.

Sources: 21, each with the date we checked it
  • canirun.ai model pageschecked Sept. 15, 2026Minimum and recommended memory per model at roughly 4-bit (Q4_K_M). Estimates, not measurements. Each model page was opened individually on Sept. 15, 2026.
  • llama.cpp CUDA scoreboardchecked Sept. 15, 2026LLaMA 7B Q4_0 llama-bench tg128 and pp512, flash attention off. One contributor's system and build per row.
  • llama.cpp gpt-oss guidechecked Sept. 15, 2026gpt-oss-20b llama-bench figures with flash attention on, prompt-processing at depth, and the --n-cpu-moe offload recipe.
  • llama.cpp Apple Silicon threadchecked Sept. 15, 2026Measured LLaMA 7B Q4_0 pp512 and tg128 by Apple chip.
  • GPUDojo listing trackerchecked Sept. 15, 2026Lowest current store listings and card specifications by model, tracked at 16:02 UTC on Sept. 15, 2026 and reread at 00:2x UTC Sept. 16. Listing prices are before tax and shipping, and are asks rather than completed sales. Memory bandwidth is the card specification.
  • GetPCParts eBay sold priceschecked Sept. 12, 2026eBay sold-listing price for the RTX 3090, as of Sept. 12, 2026. A completed sale, not an ask.
  • ResalePrices active eBay askschecked Sept. 15, 2026Fair asking range across 291 active eBay listings, Sept. 15, 2026. Asks are not sales.
  • LLM Garage dual-3090 runchecked Sept. 15, 2026gpt-oss-120b Q4_K_M on two RTX 3090s with --n-cpu-moe 26 and 64GB of DDR4, Ryzen 7 5800X, 38,000-token context, January 2026.
  • Compute Market multi-GPU guidechecked Sept. 15, 2026Community-derived estimate of 14-16 t/s for a 70B model at Q4_K_M on two 3090s, and a 1,000W-minimum supply recommendation. Updated July 19, 2026. An estimate, not a measurement.
  • Nvidia card specificationschecked Sept. 15, 2026Rated card power, recommended system supply wattage, connector count and Founders Edition dimensions. Partner boards differ.
  • Framework Desktop store pagechecked Sept. 16, 2026Price, included configuration and stock state on Framework's own store, reread Sept. 16, 2026. The DIY Edition price is the system without a drive; Framework's cheapest 500GB option adds $135.
  • GMKtec store, EVO-X2 configuratorchecked Sept. 16, 2026Per-configuration U.S. price and in-stock state read out of the maker's own store on Sept. 16, 2026: 128GB + 1TB at $3,499.99 and 128GB + 2TB at $3,649.99, both available.
  • Apple U.S. store priceschecked Sept. 2, 2026Apple's U.S. list prices as observed Aug. 25 and Sept. 2, 2026.
  • Hard Bargains used-Mac crawlerchecked Sept. 2, 2026Median asking prices from our own crawler. Asking prices mix used, open-box and refurbished condition and are not completed sales.
  • Daring Fireball August configuration rundownchecked Sept. 15, 2026Rounded new-Mac configuration prices published in August 2026.
  • MacStories early Mac Studio AI benchmarkschecked Sept. 15, 2026Qwen3-235B-A22B measured in LM Studio on a 512GB M3 Ultra.
  • Hostbor M3 Ultra testchecked Sept. 15, 2026DeepSeek R1 671B at 4-bit measured on a 512GB M3 Ultra.
  • Amazon product-page checkchecked Sept. 16, 2026Each exact ASIN below was opened Sept. 15, 2026 and reopened Sept. 16 to confirm it resolves to a live product detail page whose title names the stated variant. We checked the title only, not price, seller or stock, and no Amazon price or availability appears on this page; every price here is dated and comes from a store we could read directly.
  • Tom's Hardware used-GPU guidancechecked Sept. 15, 2026What to ask a used-card seller and how to test a card on arrival.
  • Hard Bargains memory comparisonchecked Aug. 26, 2026Our own August comparison. Memory bandwidth figures there are manufacturer specifications; the DGX Spark price is as reported in August and was not rechecked.
  • llama.cpp discussion of the Metal memory splitchecked Sept. 16, 2026Developers report that macOS withholds a third of unified memory from the GPU at 32GB and below and a quarter above it, which leaves a 16GB Mac about 10.7GB to load a model into. Developer reports, not an Apple specification.