A100 40G — what LLMs can it run?
91 of 110 indexed models fit comfortably in 40GB at 4K context, each at the highest-quality quant that still leaves headroom.
The short answer for A100 40G
40 GB, 1,555 GB/s HBM2 (PCIe figure. The SXM module is 1,555 GB/s as well.). 91 of 110 models in this index fit comfortably at 4K context; 94 load at all.
- Biggest that fits
- Nemotron 3.5 Bolt 74B-A9B — 23.1 GB at IQ2_XXS — 17.0 GB spare at 4K, so longer context comes out of a thin margin.
- Speed ceiling
- Llama 3.1 8B Instruct at Q4_K_M reads 4.6 GB of weights per token, so 1,555 GB/s puts a hard ceiling near 337 tok/s. That is arithmetic on two published numbers, not a benchmark — real throughput lands below it. No run on this card has been measured here, so there is nothing to compare the ceiling against.
- The wall you will hit
- Capacity. 91 of 110 models clear 40 GB comfortably at 4K context, and the ones that do not are short by whole gigabytes rather than by a rounding error — at 1,555 GB/s the memory system is not what holds this card back. Moving up a size class is what changes the list, not a different quant.
70B+ · 4
| Model | Quant | Est. VRAM | Headroom | tok/s |
|---|---|---|---|---|
| Nemotron 3.5 Bolt 74B-A9B74B-A9B | IQ2_XXS | 23.05 GB | +17 GB | ≤611Ceiling |
| Qwen2.5 72B Instruct72B | EXL2 3.5bpw | 34.61 GB | +5.4 GB | ≤51Ceiling |
| Llama 3.1 70B Instruct70B | EXL2 3.5bpw | 33.65 GB | +6.4 GB | ≤53Ceiling |
| Jamba 1.5 Mini52B-A12B | Q4_K_M | 33.01 GB | +7 GB | ≤225Ceiling |
32B · 30
| Model | Quant | Est. VRAM | Headroom | tok/s |
|---|---|---|---|---|
| Kimi Linear 48B-A3B Instruct48B-A3B | Q4_K_M | 30.44 GB | +9.6 GB | ≤900Ceiling |
| Mixtral 8x7B Instruct47B MoE | Q4_K_M | 30.13 GB | +9.9 GB | ≤209Ceiling |
| Seed-OSS 36B Instruct36B | Q5_K_M | 27.81 GB | +12.2 GB | ≤64Ceiling |
| Qwen3.6 35B-A3B35B-A3B | Q4_K_M | 23.38 GB | +16.6 GB | ≤871Ceiling |
| Command R 35B35B | Q4_K_M | 22.86 GB | +17.1 GB | ≤77Ceiling |
| Yi 1.5 34B Chat34B | Q4_K_M | 22.82 GB | +17.2 GB | ≤78Ceiling |
| Qwen3 32B Instruct32B | Q4_K_M | 21.85 GB | +18.2 GB | ≤82Ceiling |
| Qwen2.5 32B Instruct32B | Q4_K_M | 21.69 GB | +18.3 GB | ≤83Ceiling |
| Qwen2.5-Coder 32B Instruct32B | Q4_K_M | 21.69 GB | +18.3 GB | ≤83Ceiling |
| DeepSeek-R1-Distill-Qwen-32B32B | Q4_K_M | 21.69 GB | +18.3 GB | ≤83Ceiling |
| Nemotron 3.5 Lightning 30B-A3B30B-A3B | Q8_0 | 35.14 GB | +4.9 GB | ≤513Ceiling |
| Gemma 4 31B IT31B | Q4_K_M | 21.08 GB | +18.9 GB | ≤88Ceiling |
| Qwen3 30B-A3B Instruct30B-A3B | Q5_K_M | 23.04 GB | +17 GB | ≤769Ceiling |
| Qwen3-Coder 30B-A3B Instruct30B-A3B | Q5_K_M | 23.04 GB | +17 GB | ≤769Ceiling |
| Qwen3-VL 30B-A3B Instruct30B-A3B | Q8_0 | 34.28 GB | +5.7 GB | ≤514Ceiling |
| GLM-4.7-Flash30B-A3B | Q8_0 | 33.47 GB | +6.5 GB | ≤514Ceiling |
| Granite 4.2 30B30B | Q8_0 | 33.61 GB | +6.4 GB | ≤53Ceiling |
| Muse Glimmer 30B30B | Q8_0 | 31.08 GB | +8.9 GB | ≤55Ceiling |
| Qwen3.8 27B27B | Q4_K_M | 17.47 GB | +22.5 GB | ≤99Ceiling |
| Qwen3.6 27B27B | Q8_0 | 30.71 GB | +9.3 GB | ≤56Ceiling |
| Gemma 3 27B IT27B | Q4_K_M | 18.37 GB | +21.6 GB | ≤99Ceiling |
| Gemma 2 27B Instruct27B | Q5_K_M | 21.76 GB | +18.2 GB | ≤85Ceiling |
| Gemma 4 26B-A4B IT26B-A4B | Q8_0 | 28.42 GB | +11.6 GB | ≤385Ceiling |
| Mistral Small 24B Instruct24B | EXL2 4.65bpw | 15.26 GB | +24.7 GB | ≤117Ceiling |
| LFM2.5 24B-A2B24B-A2B | Q8_0 | 26.73 GB | +13.3 GB | ≤628Ceiling |
| Devstral Small 1.1 24B24B | Q6_K | 20.91 GB | +19.1 GB | ≤85Ceiling |
| Magistral Small 1.2 24B24B | Q6_K | 20.91 GB | +19.1 GB | ≤85Ceiling |
| Codestral 22B22B | Q4_K_M | 15.03 GB | +25 GB | ≤122Ceiling |
| ERNIE 4.5 21B-A3B21B-A3B | Q8_0 | 24.44 GB | +15.6 GB | ≤514Ceiling |
| GPT-OSS 20B21B MoE | MXFP4 | 12.76 GB | +27.2 GB | ≤786Ceiling |
14B · 18
| Model | Quant | Est. VRAM | Headroom | tok/s |
|---|---|---|---|---|
| InternLM2 20B Chat20B | Q5_K_M | 15.59 GB | +24.4 GB | ≤116Ceiling |
| DeepSeek-Coder-V2-Lite Instruct16B-A2.4B | Q8_0 | 17.56 GB | +22.4 GB | ≤642Ceiling |
| DeepSeek-V2-Lite Chat16B-A2.4B | Q4_K_M | 10.08 GB | +29.9 GB | ≤1125Ceiling |
| StarCoder2 15B15B | Q4_K_M | 10.19 GB | +29.8 GB | ≤180Ceiling |
| Qwen3 14B Instruct14B | Q5_K_M | 11.65 GB | +28.4 GB | ≤156Ceiling |
| Qwen2.5 14B Instruct14B | Q5_K_M | 11.73 GB | +28.3 GB | ≤157Ceiling |
| DeepSeek-R1-Distill-Qwen-14B14B | EXL2 4.65bpw | 9.75 GB | +30.3 GB | ≤192Ceiling |
| Phi-4 14B14B | Q5_K_M | 11.77 GB | +28.2 GB | ≤157Ceiling |
| Phi-3 Medium 14B Instruct14B | Q6_K | 12.86 GB | +27.1 GB | ≤143Ceiling |
| Mistral Nemo 12B Instruct12B | Q6_K | 11.14 GB | +28.9 GB | ≤164Ceiling |
| Gemma 3 12B IT12B | Q5_K_M | 9.84 GB | +30.2 GB | ≤189Ceiling |
| Mellum2.1 12B-A2.5B Thinking12B-A2.5B | Q8_0 | 13.6 GB | +26.4 GB | ≤617Ceiling |
| Stable LM 2 12B Chat12B | Q4_K_M | 8.35 GB | +31.7 GB | ≤223Ceiling |
| Llama 3.2 11B Vision Instruct11B | Q8_0 | 12.9 GB | +27.1 GB | ≤140Ceiling |
| Solar 10.7B Instruct11B | Q4_K_M | 7.6 GB | +32.4 GB | ≤252Ceiling |
| Falcon 3 10B Instruct10B | Q4_K_M | 7.28 GB | +32.7 GB | ≤260Ceiling |
| Gemma 2 9B Instruct9B | Q8_0 | 11.7 GB | +28.3 GB | ≤167Ceiling |
| Qwen3.5 9B9B | Q8_0 | 10.37 GB | +29.6 GB | ≤167Ceiling |
7B · 25
| Model | Quant | Est. VRAM | Headroom | tok/s |
|---|---|---|---|---|
| GLM-4-9B-Chat9B | Q8_0 | 10.16 GB | +29.8 GB | ≤171Ceiling |
| Qwen3-VL 8B Instruct8B | Q8_0 | 10.39 GB | +29.6 GB | ≤175Ceiling |
| Granite 4.2 8B8B | Q8_0 | 10.45 GB | +29.6 GB | ≤175Ceiling |
| Ministral 3 8B Instruct8B | Q8_0 | 10.02 GB | +30 GB | ≤181Ceiling |
| Qwen2-VL 7B Instruct7B | Q4_K_M | 5.49 GB | +34.5 GB | ≤326Ceiling |
| Granite 3.1 8B Instruct8B | Q4_K_M | 5.88 GB | +34.1 GB | ≤329Ceiling |
| Qwen3 8B Instruct8B | Q6_K | 7.64 GB | +32.4 GB | ≤244Ceiling |
| Llama 3.1 8B Instruct8B | Q8_0 | 9.47 GB | +30.5 GB | ≤192Ceiling |
| Nous Hermes 3 Llama 3.1 8B8B | EXL2 4.65bpw | 5.43 GB | +34.6 GB | ≤351Ceiling |
| Aya 23 8B8B | Q4_K_M | 5.64 GB | +34.4 GB | ≤337Ceiling |
| OpenChat 3.6 8B8B | EXL2 4.65bpw | 5.43 GB | +34.6 GB | ≤351Ceiling |
| DeepSeek-R1-Distill-Llama-8B8B | Q5_K_M | 6.51 GB | +33.5 GB | ≤287Ceiling |
| InternLM2 7B Chat7B | Q4_K_M | 5.45 GB | +34.6 GB | ≤349Ceiling |
| Qwen2.5 7B Instruct7B | Q6_K | 6.77 GB | +33.2 GB | ≤262Ceiling |
| Qwen2.5-Coder 7B Instruct7B | EXL2 4.65bpw | 4.87 GB | +35.1 GB | ≤369Ceiling |
| DeepSeek-R1-Distill-Qwen-7B7B | EXL2 4.65bpw | 4.87 GB | +35.1 GB | ≤369Ceiling |
| Gemma 4 E4B ITE4B | Q8_0 | 8.46 GB | +31.5 GB | ≤205Ceiling |
| OLMo 2 7B Instruct7B | Q8_0 | 10.3 GB | +29.7 GB | ≤211Ceiling |
| OLMo 3 7B Instruct7B | Q8_0 | 10.3 GB | +29.7 GB | ≤211Ceiling |
| WizardLM-2 7B7B | Q4_K_M | 5.14 GB | +34.9 GB | ≤373Ceiling |
| Mistral 7B Instruct v0.37B | Q6_K | 6.75 GB | +33.3 GB | ≤276Ceiling |
| Zephyr 7B Beta7B | Q6_K | 6.75 GB | +33.3 GB | ≤276Ceiling |
| Gemma 4 E2B ITE2B | Q8_0 | 5.2 GB | +34.8 GB | ≤332Ceiling |
| Qwen3.5 4B4B | Q8_0 | 4.97 GB | +35 GB | ≤354Ceiling |
| Gemma 3 4B IT4B | Q8_0 | 5.05 GB | +35 GB | ≤358Ceiling |
≤3B · 14
| Model | Quant | Est. VRAM | Headroom | tok/s |
|---|---|---|---|---|
| Qwen3 4B Instruct4B | Q6_K | 4.05 GB | +36 GB | ≤498Ceiling |
| Phi-4 Mini Instruct3.8B | Q8_0 | 4.81 GB | +35.2 GB | ≤401Ceiling |
| Phi-3.5 Mini Instruct3.8B | Q8_0 | 5.89 GB | +34.1 GB | ≤403Ceiling |
| Granite 4.2 3B3B | Q8_0 | 4.41 GB | +35.6 GB | ≤421Ceiling |
| Llama 3.2 3B Instruct3B | Q8_0 | 4.05 GB | +36 GB | ≤480Ceiling |
| Qwen2.5 3B Instruct3B | Q8_0 | 3.59 GB | +36.4 GB | ≤498Ceiling |
| SmolLM3 3B3B | Q8_0 | 3.73 GB | +36.3 GB | ≤500Ceiling |
| Gemma 2 2B Instruct2B | Q8_0 | 3.34 GB | +36.7 GB | ≤591Ceiling |
| Qwen3.5 2B2B | Q8_0 | 2.23 GB | +37.8 GB | ≤785Ceiling |
| Qwen3 1.7B Instruct1.7B | Q8_0 | 2.39 GB | +37.6 GB | ≤894Ceiling |
| Qwen2.5 1.5B Instruct1.5B | Q8_0 | 1.83 GB | +38.2 GB | ≤1003Ceiling |
| Llama 3.2 1B Instruct1B | Q8_0 | 1.51 GB | +38.5 GB | ≤1244Ceiling |
| Qwen3.5 0.8B0.8B | Q8_0 | 0.94 GB | +39.1 GB | ≤1920Ceiling |
| Qwen2.5 0.5B Instruct0.5B | Q8_0 | 0.6 GB | +39.4 GB | ≤3173Ceiling |
How this list is built
Each row is the lowest-perplexity-loss quant of that model whose estimated total — weights plus KV cache at 4K context plus activation buffer — uses at most 88% of the card. That is the calculator's "green" threshold, so every row here has real headroom rather than only just fitting. Raise the context length and the list shortens; the calculator lets you check any combination directly. The speed column is this card: a measured run, a range scaled from this card's own Llama 2 7B benchmark, or a bandwidth ceiling. A figure that only exists for the RTX 4090 is labelled as that card.
A further 3 models load but with no headroom to spare (up to 105% of VRAM) — the Quant Hub’s GPU chips count those too, which is why its number is higher.
Measured on this card
No benchmark runs in this index were recorded on a A100 40G. Every figure on this page is calculated from the model architecture and the quant level — treat them as estimates, not measurements.
Stepping up
A L40S 48G (48GB) fits 3 more of the indexed models than this card. L40S 48G →
Common questions
What is the best local LLM for an A100 40G?
For everyday use, Nemotron 3.5 Bolt 74B-A9B at IQ2_XXS — about 23.1 GB of the card's 40 GB at 4K context, leaving 17.0 GB for a longer window. If you want the largest thing that will load, that is Nemotron 3.5 Bolt 74B-A9B at IQ2_XXS (23.1 GB, 17.0 GB spare). "Best" here means best fit for the memory budget — this index does not run task benchmarks, so it cannot tell you which model is smarter.
Can an A100 40G run Llama 3.2 90B Vision Instruct?
No. Its smallest build here, Q3_K_M, needs about 46.2 GB and this card has 40.0 GB usable — short by 6.2 GB before any context beyond 4K. The largest model this card does clear is Nemotron 3.5 Bolt 74B-A9B.
How many tokens per second does an A100 40G do on an 8B model at Q4?
This index has no measured run on this card, so it does not publish a figure. What can be stated from specifications: generating a token requires reading every weight once, Llama 3.1 8B Instruct at Q4_K_M is 4.6 GB of weights, and this card moves 1,555 GB/s — a ceiling near 337 tok/s. Batch size, context length, the runtime and how much of the model sits in cache all take you below it.
A100 40G or L40S 48G for local LLMs?
They do not hold the same set: A100 40G fits 91 and L40S 48G 94 of 110 at 4K. Speed is the other difference — 1,555 GB/s against 864 GB/s, a 1.8× gap in how fast the weights can be read, which is what token generation is bound by. The A100 40G generates faster on any model both can hold.
91 of 110 indexed models fit comfortably in 40GB at 4K context, each at the highest-quality quant that still leaves headroom.
This page's figures change when the model or the runtime does.Last updated 2026-10-18 RSS → /feed.xml