About Mistral Small 4
Mistral Small 4 collapses what were three Mistral product lines into one open release: its model card describes a hybrid that acts as instruction model, reasoning model and coding specialist in a single Apache 2.0-licensed system, with a reasoning-effort control to move between modes. Architecturally it is a mixture-of-experts with a notably small active set per token drawn from a large expert pool, plus image input and strong system-prompt adherence.
It succeeds the earlier compact Mistral and Magistral models as the company's recommended small tier for chat assistants, coding and agentic tasks.
Mistral Small 4 is Mistral's 119B-parameter model released in March 2026, with a 256K-token context window. It is a mixture-of-experts model: all 119B parameters must sit in memory, but only ~6.5B are active per token, which is what makes it fast for its size. It uses a classic dense-attention design whose KV cache grows linearly with context: its KV cache is about 4.8 GB at an 8K context, 77.3 GB at 128K, and 154.6 GB at the full 256K window (FP16 cache).
Don't let 'Small' set expectations — the branding describes the active compute, while VRAM planning follows the full expert pool in the tables above, which lands it in workstation and large-unified-memory territory rather than gaming-card range at quality quantisations. Like its bigger sibling it ships FP8-native, so the primary release is compact for what it is, and community GGUF quantisations extend the reach further down. Its conventional attention means the KV cache does grow with context in the classic way; the calculator above models that, and it is worth checking before assuming the full window is usable on marginal hardware.
Mistral ships FP8-quantised weights as the primary release; community GGUF quants (Q4–Q6) are what most people run locally. Totals above include the KV cache and a realistic framework overhead, so they are what you should expect to see in practice rather than just the download size. Weight sizes are calibrated against real GGUF files — see the methodology.
Frequently asked questions
How much VRAM does Mistral Small 4 need?
At Q4_K_M with an 8K context, Mistral Small 4 needs about 86 GB (weights 73 GB + KV cache + overhead). The smallest common hardware that fits is a RTX PRO 6000 Blackwell.
Can an RTX 4090 (24GB) run Mistral Small 4?
Not fully in VRAM. Mistral Small 4 needs about 86 GB even at Q4_K_M, so a 24 GB card would have to offload layers to system RAM at a large speed penalty.
Can a Mac run Mistral Small 4?
Yes — Apple Silicon with 128 GB of unified memory or more (macOS lets the GPU use ~75% of it, ~96 GB) runs Mistral Small 4 at Q4_K_M.
How much VRAM does Mistral Small 4 need at its full 256K context?
The KV cache grows with context: about 4.8 GB at 8K tokens versus 154.6 GB at the full 256K window (FP16 cache). Add that to the weights (73 GB at Q4_K_M) plus overhead — long contexts can cost more than a whole quantisation step.
Related
VRAM calculator for any model ·Token counter · Mistral Small 4 API pricing
Similar-size local models: GPT-OSS 120B (117B) · Llama 4 Scout (109B) · Llama 3.3 70B (70.6B) · DeepSeek V4 Flash (284B) · Qwen3.5 397B-A17B (397B)
Last updated 2026-09-03. Architecture figures from the model's published config.json; see the methodology.