llmrun

NVIDIA · Ampere

Best AI Models for NVIDIA A100 40GB PCIe

Runs 27–31B and 35B MoE models well; 70B only just fits.

40 GB HBM2e1555 GB/s312 TFLOPS FP16250 WAmpere 2020

What LLMs can the A100 40GB PCIe run?

#ModelSpeedllmrun ScoreQuant · VRAM
1
Qwen3.8 27BAlibabaHow to run →
Q4_K_M · 17.9 GB · 58.0 t/s
58.0 tok/sQ4_K_M · 17.9 GB
2
Qwen3.6 27BAlibaba
Q4_K_M · 17.9 GB · 58.0 t/s
58.0 tok/sQ4_K_M · 17.9 GB
3
Gemma 4 31B ITGoogle
Q4_K_M · 21.7 GB · 49.6 t/s
49.6 tok/sQ4_K_M · 21.7 GB
4
Gemma 4 26B A4B ITGoogle
Q4_K_M · 16.5 GB · 229 t/s
229 tok/sQ4_K_M · 16.5 GB
5
NVIDIA Nemotron 3.5 Lightning 30B A3B BF16NVIDIA
Q4_K_M · 19.3 GB · 233 t/s
233 tok/sQ4_K_M · 19.3 GB
See all 356 models on the A100 40GB PCIe →Best quant that fits · speeds estimated on this card

About the A100 40GB PCIe

The A100 40GB PCIe runs 27–31B and 35B MoE models well. Ampere generation, released 2020.

GenerationAmperereleased 2020

A100 40GB PCIe strengths and weaknesses

Strengths

  • Runs 70B models: Llama 3.3 70B Instruct at 28.9 tok/s in 39.6 GB
  • Room for long context: Qwen3.8 27B at Q4 uses 17.9 GB, leaving 22.1 GB free
  • Reads long prompts fast: A 4K-token prompt on Qwen3.8 27B starts answering in 1.6 s

Weaknesses

  • 120B MoE models don't fit: GPT OSS 120B needs 70.8 GB at Q4 vs 40 GB here
  • No FP8 or NVFP4: Ampere predates both; use GGUF, AWQ or GPTQ builds

A100 40GB PCIe tokens per second by model size

Output speed depends on memory bandwidth (1555 GB/s). Grey rows need more than 40 GB.

  1. 9BQwen3.5 9B · Q4_K_M · 6.6 GB
    159tok/s
  2. 14BQwen3 14B · Q4_K_M · 11.9 GB
    106tok/s
  3. 27–31BQwen3.8 27B · Q4_K_M · 17.9 GB
    58.0tok/s
  4. 35B MoEQwen3.6 35B A3B · Q4_K_M · 22.0 GB
    197tok/s
  5. 70BLlama 3.3 70B Instruct · IQ3_M · 39.6 GB tight · IQ3_M
    28.9tok/s
  6. 120B MoEGPT OSS 120B · not at 4-bit
    needs 70.8 GB

Estimated speeds: Q4_K_M, 4K context, one user. How we estimate

All 1202 models on the A100 40GB PCIe

Featured only

Best quant that fits, ordered by llmrun Score. Fit grades how well it runs here, S best. How ranking works

LLM models compatible with NVIDIA A100 40GB PCIe — ranked by Score, recency and Fit
ModelVRAMScoreFit
MN VelvetCafe RP 12B
Q4_K_M·125.2 t/s
11.0 GBS95, Runs great
Llamatron 8B V1
Q4_K_M·187.5 t/s
7.3 GBS99, Runs great
TeleOCR
Q4_K_M·795.9 t/s
2.1 GBS100, Runs great
Recursive Language Model 198M
BF16·2297.2 t/s
0.4 GBS100, Runs great
VulnLLM R 7B
Q4_K_M·202.6 t/s
5.8 GBS100, Runs great
Huihui Qwen3 Coder 30B A3B Instruct Abliterated
Q4_K_M·214.0 t/s
19.4 GBS100, Runs great
Gemma 4 E4B Luchador
BF16·61.2 t/s
16.8 GBA84, Runs well
Browsesafe
Q4_K_M·214.0 t/s
19.4 GBS100, Runs great
SmolLM2 70M
Q4_K_M·2659.9 t/s
0.5 GBS100, Runs great
Qwen3 0.6B Heretic Abliterated Uncensored
Q4_K_M·1295.8 t/s
1.6 GBS100, Runs great
LightOnOCR 2 1B
Q4_K_M·990.9 t/s
1.8 GBS100, Runs great
Tiny Random Gemma2ForCausalLM
Q4_K_M·3260.5 t/s
0.3 GBS100, Runs great
AFM 4.5B
Q4_K_M·313.9 t/s
4.3 GBS100, Runs great
Qwen3.5 2B Text Only
BF16·243.0 t/s
4.3 GBS100, Runs great
Qwen3 8B 64k Context 2X Josiefied Uncensored
Q4_K_M·183.1 t/s
7.6 GBS99, Runs great

A100 40GB PCIe specs, power and PSU

Specs

Architecture
Ampere
VRAM
40 GB HBM2e
Memory bandwidth
1555.0 GB/s
FP16 (dense)
312.00 TFLOPS
CUDA cores
6,912
Tensor cores
432
Compute capability
8.0
TDP
250 W
Released
2020-05-14

A100 40GB PCIe FAQ

What is the largest model the A100 40GB PCIe can run?

Qwen3.6 35B A3B (36.0B) is the largest at 4-bit or better: at Q4_K_M it uses 22.0 GB of 40 GB, and runs at about 197 tok/s.

Can the A100 40GB PCIe run Qwen3.8 27B?

Yes. Qwen3.8 27B fits at Q4_K_M, using 17.9 GB of 40 GB, and it is the best model that runs well on this card.

Can the A100 40GB PCIe run 120B MoE models?

No. GPT OSS 120B needs 70.8 GB at Q4, more than the 40 GB on this card.

How many tokens per second does the A100 40GB PCIe give?

About 58.0 tok/s on Qwen3.8 27B at Q4_K_M, with the first token after 1.6 s on a 4K-token prompt.

What power supply does the A100 40GB PCIe need?

Plan a 500 W PSU (2× the 250 W TDP).

Is 40 GB enough for coding and agents?

Yes for coding: Qwen3.8 27B, the best of the 10 coding models that fit, runs at about 58 tok/s. For agents, not yet: no agent-scored model fits in less than 96 GB.

Related

Best coding LLMs24 GB VRAM48 GB VRAMGPUs with 40 GB or more