llmrun

NVIDIA · Ampere · 48 GB VRAM tier

Best AI Models for NVIDIA A40 (48 GB)

Runs 27–31B and 35B MoE models well; 70B barely fits and is slow (11 tok/s).

48 GB GDDR6696 GB/s150 TFLOPS FP16300 WAmpere 2020

What LLMs can the A40 run?

#ModelSpeedllmrun ScoreQuant · VRAM
1
Qwen3.8 27BAlibabaHow to run →
Q4_K_M · 17.9 GB · 26.0 t/s
26.0 tok/sQ4_K_M · 17.9 GB
2
Qwen3.6 27BAlibaba
Q4_K_M · 17.9 GB · 26.0 t/s
26.0 tok/sQ4_K_M · 17.9 GB
3
Gemma 4 31B ITGoogle
Q4_K_M · 21.7 GB · 22.2 t/s
22.2 tok/sQ4_K_M · 21.7 GB
4
Gemma 4 26B A4B ITGoogle
Q4_K_M · 16.5 GB · 130 t/s
130 tok/sQ4_K_M · 16.5 GB
5
NVIDIA Nemotron 3.5 Lightning 30B A3B BF16NVIDIA
Q4_K_M · 19.3 GB · 165 t/s
165 tok/sQ4_K_M · 19.3 GB
See all 368 models on the A40 →Best quant that fits · speeds estimated on this card

About the A40

The A40 runs 27–31B and 35B MoE models well. Ampere generation, released 2020.

GenerationAmperereleased 2020

A40 strengths and weaknesses

Strengths

  • Runs 35B MoE models: Qwen3.6 35B A3B at 116 tok/s in 22 GB
  • Room for long context: Qwen3.8 27B at Q4 uses 17.9 GB, leaving 30.1 GB free

Weaknesses

  • 120B MoE models don't fit: GPT OSS 120B needs 70.8 GB at Q4 vs 48 GB here
  • No FP8 or NVFP4: Ampere predates both; use GGUF, AWQ or GPTQ builds

A40 tokens per second by model size

Output speed depends on memory bandwidth (696 GB/s). Grey rows need more than 48 GB.

  1. 9BQwen3.5 9B · Q4_K_M · 6.6 GB
    71.1tok/s
  2. 14BQwen3 14B · Q4_K_M · 11.9 GB
    47.6tok/s
  3. 27–31BQwen3.8 27B · Q4_K_M · 17.9 GB
    26.0tok/s
  4. 35B MoEQwen3.6 35B A3B · Q4_K_M · 22.0 GB
    116tok/s
  5. 70BLlama 3.3 70B Instruct · IQ4_XS · 46.4 GB tight · IQ4_XS
    10.8tok/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 1216 models on the A40

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 A40 — ranked by Score, recency and Fit
ModelVRAMScoreFit
Dots.ocr
Q4_K_M·207.5 t/s
2.6 GBS100, Runs great
Tiny MistralForCausalLM 0.2
Q4_K_M·1508.0 t/s
0.5 GBS100, Runs great
SmolLM3 3B ONNX
Q4_K_M·201.1 t/s
3.3 GBS100, Runs great
Tiny MistralForCausalLM 0.1
Q4_K_M·1508.0 t/s
0.5 GBS100, Runs great
Falcon H1 0.5B Instruct
Q4_K_M·655.7 t/s
1.2 GBS100, Runs great
Qwen3 Embedding 0.6B
Q4_K_M·587.5 t/s
1.6 GBS100, Runs great
Tower Plus 9B
Q4_K_M·70.0 t/s
7.7 GBS86, Runs great
OmniSVG1.1 8B
BF16·27.6 t/s
17.2 GBA68, Runs well
Tiny GemmaForCausalLM
Q4_K_M·1508.0 t/s
0.4 GBS100, Runs great
Lucid V1 Nemo
Q4_K_M·56.1 t/s
11.0 GBA82, Runs well
Tiny FalconMambaForCausalLM
BF16·45240.0 t/s
0.2 GBS100, Runs great
Tiny LlamaForCausalLM 3
Q4_K_M·1508.0 t/s
0.4 GBS100, Runs great
Qwen3.5 9B Humanize DPO Round2
BF16·22.8 t/s
22.4 GBB63, Decent
Tiny LlamaForCausalLM 3.1
Q4_K_M·1508.0 t/s
0.6 GBS100, Runs great
T5gemma B B Ul2 IT
BF16·348.0 t/s
2.5 GBS100, Runs great

Rent NVIDIA A40 in the cloud

Hourly rate on demand for one A40.

  • RunPod$5+ bonus creditLowest price

    from$0.35/hr

    Deploy

  • Vast.aiMarketplace

    from$0.71/hr

    Deploy

Cost per million tokens

Output-token cost on an A40 rented at $0.35/hr, one request at a time. Batch several users on the same card and the cost per token drops several-fold.

ModelSpeedPer 1M output
NVIDIA Nemotron 3.5 Lightning 30B A3B BF16165 tok/s$0.59$0.0068 input
Gemma 4 26B A4B IT130 tok/s$0.75$0.013 input
Qwen3.6 35B A3B116 tok/s$0.84$0.014 input
Qwen3.8 27B26 tok/s$3.74$0.080 input
Muse Glimmer 30B25 tok/s$3.94$0.086 input

A40 specs, power and PSU

Specs

Architecture
Ampere
VRAM
48 GB GDDR6
Memory bandwidth
696.0 GB/s
FP16 (dense)
149.70 TFLOPS
CUDA cores
10,752
Tensor cores
336
Compute capability
8.6
TDP
300 W
Released
2020-10-05

A40 FAQ

What is the largest model the A40 can run?

Qwen2.5 72B Instruct (72.7B) is the largest at 4-bit or better: at Q4_K_S it uses 46.6 GB of 48 GB, and runs at about 10.8 tok/s.

Can the A40 run Qwen3.8 27B?

Yes. Qwen3.8 27B fits at Q4_K_M, using 17.9 GB of 48 GB, and it is the best model that runs well on this card.

Can the A40 run 120B MoE models?

No. GPT OSS 120B needs 70.8 GB at Q4, more than the 48 GB on this card.

How many tokens per second does the A40 give?

About 26.0 tok/s on Qwen3.8 27B at Q4_K_M, with the first token after 3.4 s on a 4K-token prompt.

What power supply does the A40 need?

Plan a 600 W PSU (2× the 300 W TDP).

Is 48 GB enough for coding and agents?

Yes for coding: Qwen3.8 27B, the best of the 10 coding models that fit, runs at about 26 tok/s. For agents, not yet: no agent-scored model fits in less than 96 GB.

Related

Best coding LLMs48 GB VRAMGPUs with 48 GB or more