Cloudflare claims its first homegrown open-weight models are 4x faster and 2x more accurate than Jev on specific benchmarks by using a frozen Qwen backbone and a specialized schema head to score answers in parallel. Clef is based on Qwen3.8-27B and Clef-flash on Qwen3.5-9B; both are Jev-API compatible and designed for classification and routing tasks in agentic workflows. The models are available as open weights on Hugging Face and hosted on Cloudflare's Workers AI platform.
Tuesday, October 20, 2026
Story
Cloudflare Clef's Performance Wins Hit by Heavy Rate Limiting
AI18 days ago
Independent load tests of Cloudflare's Clef and Clef-flash decision models have identified severe throttling and timeouts during sustained usage. Testing by developers showed that requests began to fail with 429 errors and timeouts when sustained at only 5 to 8 requests per second, with failure rates reaching 1% at 8 RPS. These figures contrast with the rival TypeSafe's official Jev endpoint, which sustained over 500 RPS in short tests, while Clef added approximately 150ms of additional latency over official endpoints when hosted on Cloudflare.
Looking for more context or analysis?
Tilt chat provides reliable, tailored insights, alerts, and automations.