What it costs to run Mistral Small 4 119B (MoE)

How these numbers are produced

• Model shape comes from the model's own config.json — 36 layers, 32 attention heads, 1 KV heads.
• Weight size is parameters × effective bits-per-weight. Those constants are checked against real published quantised file sizes — median error 0.7% across 40 measurements.
• KV cache is 2 × layers × KV-heads × head-dim × context × 2 bytes. Using KV-heads rather than attention heads is what makes this correct for grouped-query attention; treating a GQA model as multi-head overstates the cache by up to 8×.
• Tokens/sec is an estimate, not a benchmark. Generation is memory-bandwidth-bound, so this is bandwidth ÷ weight-bytes derated to 75%. Real throughput depends on your runtime, batch size and kernels. We don't run our own hardware tests — see our editorial policy.
• This is a mixture-of-experts model. Generation reads only the experts active for each token (4 of 128), so throughput is scored on ~6.5B active parameters, not the full 119B. Memory is the opposite: every expert must still be resident in VRAM, so the figures above use the full weight set.MoE with MLA (DeepSeek-style latent attention): kv_lora_rank 256 + qk_rope_head_dim 64 = 320 cached numbers per token per layer, mapped as kv_heads 1 x head_dim 160 x 2 = 320 (the config nominal 32 KV heads x 128 would overstate KV ~26x). All 36 layers full attention and MoE: 128 routed experts (top-4) + 1 shared; intermediate_size 12288 is unused (no dense layers). Card: 119B total, 6.5B active, 256k context (config allows 1048576 via YaRN x128). Official weights are FP8; params = HF safetensors total incl. the 0.43B Pixtral vision encoder. HF config + card read Nov 7, 2026.
• Some memory-bandwidth figures are not yet independently verified.Where that's the case we show no tokens/sec at all rather than a number we can't stand behind.
• Rental prices are pulled hourly from provider APIs (last updated Sat, 07 Nov 2026 02:20:13 GMT). Spot/interruptible pricing can change or vanish without notice.