Cloudflare's open-weights decision model Clef scores all answers in one pass, 53ms per decision on RTX 5090
michellechen · x · 2026-10-04
Cloudflare has released Clef, a new open-weights decision model. A user benchmarked it on an RTX 5090: each decision takes just 53 ms.
The architecture is unusual:
- It never generates text — no token output at all
- It reads the input once, then scores every allowed answer in parallel, emitting a decision directly
The author shared a 2-minute explainer covering how it works, where it beats Jev, and where it doesn't. Score-based (rather than generative) architectures like this are especially interesting for low-latency classification and routing tasks.
Related event: Cloudflare Open-Sources Decision Models Clef, Now on Ollama(3 posts)→
More from Infra
- NVIDIA AIPerf docs go live: a package for performance-testing AI models — TheZachMueller · 2026-10-04
- One prompt freed 39.7 GB: Claude Code + ccmd MCP safely cleans dev caches — julsimon · 2026-10-04
- Community pushes llm-inference-bench as standard for local LLM inference speed measurement — TheZachMueller · 2026-10-04
- From one 3090 to 20 DGX Sparks: a home local-LLM cluster epic, 2.8T Kimi K3 at 20 t/s — ciprianveg · 2026-10-04
- Rural Queensland faces a $31B Anthropic datacentre, and locals aren't happy — nordicinst · 2026-10-04
- America's data center fight previews what's coming for the rest of the world — pstAsiatech · 2026-10-04