Clef-flash
Cloudflare's lightweight 9B sibling to Clef, built for sub-40ms decision-model inference.
Overview
Released alongside Clef on October 1, 2026, Clef-flash is post-trained on a frozen Qwen3.5-9B backbone using the same rank-256 low-rank adapter approach, trading some capacity for speed. Cloudflare reports a median latency of 38.8ms across 43 internal benchmarks versus 209.3ms for full Clef and 524.1ms for TypeSafe's Jev, with sample results including 98.76% exact-match on BFCL and 93.11% accuracy on API-Bank. It shares Clef's 64k context window, vision encoder, and typed-output format, and is open-sourced on Hugging Face under Apache 2.0.
This is a sourced directory entry. Features and capabilities are reported in the linked sources; this entry does not document independent Aymorphic testing.
Key features
- Post-trained on frozen Qwen3.5-9B backbone with rank-256 low-rank adapters
- Median latency of 38.8ms across 43 benchmarks, the fastest of the Clef/Jev decision-model cohort
- 98.76% exact-match on BFCL, 93.11% accuracy on API-Bank in vendor benchmarks
- Shares Clef's 64k context window and vision encoder for image/video classification
- Open-sourced on Hugging Face under Apache 2.0; hosted on Cloudflare Workers AI
Use cases
- Latency-sensitive routing and classification inside agent loops
- High-throughput, low-cost decision scoring where full Clef's accuracy isn't required
- Edge/Workers AI deployments needing sub-40ms responses
Pricing
Open-weight model, free to download under Apache 2.0; Workers AI hosted usage pricing not published in launch coverage.
Pricing reflects the source record and may change. Confirm current terms with the provider.
Dates and sources
- Released
- 1 Oct 2026
- Record updated
- 4 Oct 2026
- Record last checked
- 4 Oct 2026