Mistral Nemo 12B VRAM Requirements

Mistral Nemo 12B needs about 9.0 GB of VRAM at Q4_K_M with an 8K context: 6.9 GB for weights, 1.3 GB for the KV cache and 0.8 GB of runtime overhead. At Q8_0 it needs 14.2 GB and at full FP16 precision 24.8 GB. The smallest GPU that runs it fully is the Intel Arc B570 10GB.

By Mistral AI & NVIDIA · Updated October 14, 2026 · Values from config.json on Hugging Face

9.0 GB VRAM

Mistral Nemo 12B at Q4_K_M with 8K context and the KV cache in FP16.

  • Weights 6.9 GB
  • KV cache 1.3 GB
  • Runtime overhead 0.8 GB

Which GPUs can run it

Mistral Nemo 12B VRAM by quantization

Total memory with an FP16 KV cache. "Smallest tier" is the smallest standard VRAM size with a 6% safety margin at 8K.

QuantizationBits/weightWeightsTotal at 8KTotal at 32KSmallest tier
Q4_K_M4.856.9 GB9.0 GB12.7 GB10 GB
Q5_K_M5.698.1 GB10.2 GB13.9 GB12 GB
Q6_K6.599.4 GB11.4 GB15.2 GB16 GB
Q8_08.5012.1 GB14.2 GB17.9 GB16 GB
FP16 / BF1616.0022.7 GB24.8 GB28.6 GB32 GB
Q3_K_M3.915.6 GB7.6 GB11.4 GB10 GB
Q2_K3.004.3 GB6.3 GB10.1 GB8 GB

VRAM by context length

At Q4_K_M. The KV cache grows linearly with context: 0.16 GB per 1,000 tokens in FP16.

ContextKV cache (FP16)KV cache (Q8_0)Total, FP16 cacheTotal, Q8_0 cache
4K tokens0.6 GB0.3 GB8.3 GB8.0 GB
8K tokens1.3 GB0.7 GB9.0 GB8.4 GB
32K tokens5.0 GB2.7 GB12.7 GB10.4 GB
128K tokens20.0 GB10.6 GB27.7 GB18.3 GB

GPUs that can run Mistral Nemo 12B

At Q4_K_M with 8K context. "Best quant" is the highest-quality format that fits fully; "max context" is at Q4_K_M.

GPUUsable memoryStatusBest quantMax contextSpeed
Intel Arc B570 10GB10.0 GBFitsQ4_K_M10K~23 tok/s
RTX 3080 10GB10.0 GBFitsQ4_K_M10K~60 tok/s
Apple M4 16GB10.7 GBFitsQ4_K_M14K~9.6 tok/s
RTX 2080 Ti 11GB11.0 GBFitsQ5_K_M16K~49 tok/s
RTX 2060 12GB12.0 GBFitsQ5_K_M22K~27 tok/s
RTX 3060 12GB12.0 GBFitsQ5_K_M22K~29 tok/s
Radeon RX 6700 XT 12GB12.0 GBFitsQ5_K_M22K~26 tok/s
Radeon RX 7700 XT 12GB12.0 GBFitsQ5_K_M22K~29 tok/s
Intel Arc B580 12GB12.0 GBFitsQ5_K_M22K~28 tok/s
RTX 4070 12GB12.0 GBFitsQ5_K_M22K~40 tok/s
RTX 4070 Super 12GB12.0 GBFitsQ5_K_M22K~40 tok/s
RTX 4070 Ti 12GB12.0 GBFitsQ5_K_M22K~40 tok/s
RTX 5070 12GB12.0 GBFitsQ5_K_M22K~53 tok/s
RTX 3080 12GB12.0 GBFitsQ5_K_M22K~71 tok/s
RTX 3050 8GB8.0 GB84% on GPUQ2_K–~13 tok/s
RTX 3060 Ti 8GB8.0 GB84% on GPUQ2_K–~18 tok/s
RTX 3070 8GB8.0 GB84% on GPUQ2_K–~18 tok/s
RTX 3070 Ti 8GB8.0 GB84% on GPUQ2_K–~21 tok/s
RTX 4060 8GB8.0 GB84% on GPUQ2_K–~14 tok/s
RTX 4060 Ti 8GB8.0 GB84% on GPUQ2_K–~15 tok/s

About Mistral Nemo 12B

Built with NVIDIA and trained for FP8 inference without quality loss. It is popular for role-play and creative fine-tunes. With 40 layers the cache is 160 KB per token in FP16, so 128K context needs 20 GB on top of the weights.

SpecValue
Parameters12.2B
Layers40
Hidden size5120
Attention heads / KV heads32 / 8
Head dimension128
Vocabulary131,072
Max context (used here)128K tokens
KV cache per token (FP16)160 KB
Released2024-07
LicenseApache 2.0

Mistral Nemo 12B VRAM: common questions

What GPU does Mistral Nemo 12B need?

A 12 GB card runs Q4_K_M with up to about 20K context. 16 GB cards fit Q6_K or longer contexts.

How much VRAM does Mistral Nemo 12B need?

About 9.0 GB at Q4_K_M with an 8K context, 14.2 GB at Q8_0 and 24.8 GB at FP16. Longer contexts add 0.16 GB per 1,000 tokens with an FP16 cache.

What is the cheapest GPU that can run Mistral Nemo 12B?

The smallest listed device that fits it fully at Q4_K_M with 8K context is the Intel Arc B570 10GB (10.0 GB usable).

Related