A safety filter you can run yourself, with your own rules
Mistral released Shieldstral on 4 August: a 3-billion-parameter, open-weight (Apache 2.0) safety classifier that judges text and images against moderation rules you write in plain language at inference time, rather than a fixed list of harm categories baked in during training. You give it the context, a yes-or-no policy question, and the content to check — a prompt, a response, a prompt-and-response pair, or an image — and it returns a verdict. It covers 12 languages and runs on a single 16GB GPU. Mistral says it matches open guard models up to seven times its size on text safety, reporting an average F1 of 84.9 on text-safety benchmarks — level with the 20B GPT-OSS Safeguard and ahead of Qwen3Guard 8B and LlamaGuard 4 12B — and claims a new high on multimodal moderation. The weights are on Hugging Face.
For a regulated European buyer, the interesting part is where the check runs. Screening what goes into and out of an AI system has usually meant sending that content to a US-hosted moderation API — the exact data exposure the buyer was trying to avoid. A downloadable classifier that fits on one GPU and enforces your own written policies keeps the safety layer on your own infrastructure, in your own jurisdiction. One limit worth stating plainly: it classifies, it does not rewrite. You get allow or deny, not redaction — so it sits in front of a system as a gate, not as a scrubber.
Meta ships a coding model and its first agent
Meta released Muse Spark 1.2 and Muse Code on 5 August — its third Muse model in four months. Spark 1.2 is tuned for coding, with higher first-attempt accuracy and more reliable tool calling, a 1M-token context, and text-and-image input; it scores 57 on the Artificial Analysis Intelligence Index at $1.25/$4.25 per million tokens. Muse Code, shipped in beta alongside it, is a terminal agent that plans, writes and validates changes end to end. The detail an EU team should read closely is the pricing: a “contributor tier” cuts the rate to $0.10/$0.20 per million in exchange for permission to train future Meta models on your prompts and completions. It is cheap, and for anyone handling regulated data it is a governance decision, not a checkout option.
Databricks turns on a governance plane for enterprise AI
Databricks made Unity AI Gateway generally available on 4 August: one API that reaches Anthropic, OpenAI, Gemini and others, lets a company control which teams use which services, routes and budgets traffic across providers, governs the MCP servers agents connect to, and logs usage, cost and lineage in one place. This is the unglamorous plumbing — the part of an agent deployment that isn't the model — and it is where a lot of the market is now spending. It is worth watching for European teams, with one caveat: a gateway governs the request and the audit trail; it does not change where the model runs or whose law reaches it.
Quick Hits
- Rust adopts an LLM contribution policy. — On 5 August, five rust-lang/rust teams adopted a policy letting AI tools help contributors ask questions, analyse, refine and review, but not produce the final contribution — a deliberately narrow rule that covers only those teams' work, not the wider Rust ecosystem.
- First open High Bandwidth Flash standard. — SanDisk and SK hynix released the first OCP technical specification for High Bandwidth Flash at the Future of Memory and Storage summit (4–6 August) — a NAND-based memory tier sitting between HBM and SSDs, supporting up to 512GB and roughly 0.4–3.0 TB/s, aimed squarely at AI inference workloads.
- Europe's venture money is now mostly AI. — European AI startups took a record 55% of the region's venture capital in the first half of 2026, about $23B, but 73% of it went to just 38 companies raising $100M-plus rounds — concentration at the top, not breadth.