Cloudflare open-sources Clef, a 27B decision model under Apache-2.0
Calibrated choices without an LLM parsing step just went open-weight: run Clef or the 38ms Clef-Flash on your own GPU, and LTX-2.5 adds open multishot video.
Copy markdown
Decision models go open-weight
Cloudflare released Clef, a 27B Apache-2.0 multimodal model that takes a state plus a schema of typed questions and returns a calibrated probability for every allowed option in a single forward pass — no free-form text, no output parsing to babysit. It's Jev/SystemOne-compatible with a 64k context and vision input, so you can swap a proprietary decision API for weights you own, from Hugging Face or Workers AI today.
Clef-Flash: sub-40ms routing you host yourself
The 9B sibling, Clef-Flash, clocks a 38.8ms median versus 209ms for the 27B — fast and cheap enough to live in the hot path of an agent loop for classification, triage, and routing. Pair it with an LLM on Workers AI to take the action, and because the weights are Apache-2.0 the decision step carries no per-call fee.
LTX-2.5 brings open multishot video and audio
Lightricks shipped LTX-2.5, a ~22B open-weight model that generates synchronized video and audio from text, image, or video — and its new multishot mode holds character, environment, and style steady across cuts. 4K via an upscaler, plus fp8 and CPU-offload paths for lower VRAM, and free commercial use under $10M revenue, so you can cut a short film on your own GPU this weekend.