Mistral Small 4 119B — VRAM requirements
Mistral AI · 119.4B params (6.5B active, mixture-of-experts) · up to 256K context
| Quantization | VRAM at 4K context | VRAM at 256K context |
|---|---|---|
| FP16 / BF16 | 277.4 GB | 452.4 GB |
| FP8 | 140.1 GB | 315.1 GB |
| INT8 | 140.1 GB | 315.1 GB |
| INT4 | 71.4 GB | 246.5 GB |
Weights = params × bytes-per-parameter for the selected quantization.KV cache is always computed at FP16 and assumes batch size 1. A flat +15% is added on top for CUDA context/activation overhead — a rule of thumb, not a measured figure. Full methodology on the interactive calculator.
Compatible GPUs at 256K context
FP16 / BF16 (452.4 GB)
No single GPU in this site's database has enough VRAM — would need multiple GPUs (not modeled here).
FP8 (315.1 GB)
INT8 (315.1 GB)
INT4 (246.5 GB)
Sparse mixture-of-experts (128 routed experts, top-4, plus 1 shared expert per config.json). paramsB (119.40B) is the exact safetensors total from the Hugging Face API; activeParamsB (6.5B) is Mistral's own stated figure ("119B parameters, with 6.5B activated per token"). **maxContextLength uses Mistral's documented 256k, not config.json's max_position_embeddings of 1,048,576** — the card states a 256k context window in two places, so the larger config value appears to be headroom rather than a supported window. This model uses MLA-style compressed attention (config.json carries kv_lora_rank), which this site's KV-cache term does not model, so the KV figure is an overestimate; at this parameter count the weights term dominates regardless. Pulled 2026-09-19.