Ministral 3 8B

“When the constraint is 'must run on a laptop or small server,' this is the default — native vision, real reasoning, and clean Apache 2.0 in one 8B.”

For any feature constrained to a laptop or small server, Ministral 3 8B is now the default. The combination of native vision, real reasoning (AIME 78.7%), EU-language quality, and a clean Apache 2.0 license is unusual at this size. I would route premium queries to Medium 3.5 or Small 4 and let the 8B handle the long tail of simpler tasks and on-device features. For embedded EU deployments where data must stay on the customer's hardware, this is the model — full on-prem control, no license fee, modest hardware.

“Native vision plus real reasoning at 8B, fully open — it strengthens Mistral's edge-tier story against Llama and Qwen on the axes EU buyers care about.”

Strategically the 8B reinforces Mistral's edge-tier position. Native vision and credible reasoning at 8B, under clean Apache 2.0, beat Llama 4 8B and Qwen 3 8B on the EU-relevant axes (multilingual quality, vision, license clarity). It is the bridge between the on-device 3B and the more capable 14B, giving product teams a clean ladder. The differentiation is the bundle at the size, not a single benchmark. Market timing aligns with edge/on-device demand and the EU AI Act compliance tailwind; the open license maximises adoption.

“$0.15/$0.15 makes monthly forecasting boring in the best way, and self-host on a consumer GPU caps cost at infrastructure for very high volume.”

Symmetric $0.15/$0.15 pricing makes monthly forecasting trivial — request count is the only variable. Self-host under clean Apache 2.0 on a consumer GPU caps cost at infrastructure for very high volume, with no license fee (unlike Medium 3.5 or the 125B Devstral). For any cost-sensitive workload that doesn't demand flagship quality, this is the right starting point, and the reasoning variant avoids escalating to a pricier model for many queries. Strong unit economics for the embedded-AI and high-volume-chat use cases.

“Fast and 'fine' for routine tasks; the reasoning variant gives me a careful answer without paying for a bigger model, and it runs locally.”

The instruct variant is fast and reliable for routine tasks — summarisation, extraction, classification, simple agent steps. The reasoning variant is useful when I want a careful answer without a bigger model. Vision is usable for screenshot triage. Same API shape as the rest of Mistral, so swapping is trivial, and clean Apache 2.0 makes fine-tuning straightforward. Runs locally via vLLM/Ollama/llama.cpp/LM Studio at a 16GB footprint. Not a model for hard problems, but a strong default for the easy 80% with a reasoning escape hatch.

“Snappy and capable for routine tasks, occasionally rough on nuance — and EU-language quality feels native, not translated.”

Snappy and capable for routine tasks, occasionally rough on nuanced queries. Vision works on simple things. The standout is European-language quality, which feels native rather than translated. Refusal rate is moderate. With reasoning toggled on it takes longer but is noticeably more careful. Conversational warmth is mid — efficient rather than friendly. A solid "free tier" model behind consumer features, especially when on-device delivery gives instant, private responses.

“Honest published numbers and a clean license — the only thing to flag is that 8B reasoning headlines come from the reasoning variant at full effort.”

Like the 14B, this is an honest small-model launch with real published benchmarks across variants, so there's little to debunk. The fair caveats: the AIME 78.7% headline is the reasoning variant at full effort (latency/token cost), the instruct variant most apps use is less spectacular, and an 8B's ceiling on hard, broad tasks is real — it won't stand in for the 14B or Small 4 where capability matters. Vision is genuinely soft at this size. The honest claim — "fast multilingual edge model with vision and a reasoning option" — holds; just pick the right variant and don't over-extend it.