Mistral open-sources Shieldstral: 3B multimodal moderation, one GPU
Apache 2.0 weights, policies you write as plain-English questions, a 3B that grades like a 20B — plus a 20B reasoner running local at 200 tok/s.
Copy markdown
Open weights, one 16GB GPU
Mistral released Shieldstral 1.0 under Apache 2.0 — a 3B multimodal safety classifier for text and images that runs on a single 16GB GPU, with weights on Hugging Face. You can self-host prompt, response, and refusal moderation instead of metering every call through a hosted safety API.
Your safety policy is just a question
Instead of a fixed taxonomy, you pass Shieldstral your rule as a plain-language question at inference and get a calibrated yes/no probability from a single forward pass — no fine-tuning. Change the prompt to run different policies on comments, uploads, or DMs from the same model.
A 3B that grades like a 20B
It averages 84.9% F1 on text safety across 16 benchmarks — level with the 20B GPT-OSS-Safeguard — and tops multimodal at 83.8% F1, beating OmniGuard-7B (77.6%) plus larger guards like LlamaGuard-4 (12B) and ShieldGemma (9B). Smaller model, lower latency, same guardrails.
Maple-Preview: a 20B reasoner that runs local
DeepGrove open-sourced Maple-Preview (MIT), a 20B MoE with just 1B active that hits 218 tok/s on an M4 Mac mini and claims IMO-level math, in a 5.31GB checkpoint with a 131K context. A near-frontier reasoner you can run offline, no GPU bill.