Cloudflare releases Clef and Clef-flash, open-weight multimodal decision models that are Jev-API compatible
On Oct 1, 2026 Cloudflare released Clef (built on Qwen3.8-27B) and Clef-flash (built on Qwen3.5-9B), Apache-2.0 open-weight "decision models". Like TypeSafe's Jev they return probabilities for typed options instead of generating text, but they also accept images and video and have a 64k context. Both are on Hugging Face and Workers AI, where Clef costs $0.24 per 1M input tokens. Cloudflare also announced an RL fine-tuning service for them.
Key facts
- Clef: Qwen3.8-27B backbone with a vision encoder; Clef-flash: Qwen3.5-9B backbone; Apache 2.0 weights on Hugging Face (cloudflare/clef, cloudflare/clef-flash); training data not released
- Prefill-only, non-autoregressive scoring: one forward pass scores every allowed option of every question in the schema (joint schema head)
- Inputs: text, JSON, images, video; 64k context on Workers AI (Jev: 32k); strictly typed outputs, fully Jev-API compatible
- Latency across 43 eval benchmarks (Cloudflare): Clef-flash median 38.8 ms / p95 122.4 ms; Clef 209.3 / 238.6 ms; Jev 524.1 / 536.0 ms
- Selected scores (Cloudflare): Clef-flash BFCL 98.76, API-Bank 93.11; Clef ToolRet 69.19, BANKING77 94.20, CLINC150+OOS 97.43; model card lists MMLU 90.3, ARC-Challenge 97.7
- Workers AI: @cf/cloudflare/clef at $0.24 per 1M input tokens (Cloudflare docs), about 6x Jev's $0.042/M (The Register); @cf/cloudflare/clef-flash also listed
- The Register: Cloudflare claims Clef beats Jev in 3 of 4 benchmark areas; self-hosting needs at least 85 GB VRAM (Clef) or 41 GB (Clef-flash) at 64k context
- The first models trained by Cloudflare's Workers AI team (MarkTechPost, @michellechen); MarkTechPost lists Clef-flash at $0.09 per 1M input tokens on Workers AI (not confirmed from Cloudflare docs)
- Also announced: an RL fine-tuning service for decision models, starting with forward-deployed engineers and moving to self-serve (AI Gateway, Workers AI, Containers and a new Trainer component)
What happened
Two weeks after TypeSafe AI launched Jev, the hosted "System One" decision model, Cloudflare released open-weight competitors trained on Qwen backbones. They plug into the same API shape. The speed and quality comparisons with Jev are Cloudflare's own, on its own selection of benchmarks, and have not been independently checked.
Why it matters
It is the second big-company follower in two days, after AWS's Strands Decider 2B. Decision models are quickly becoming a product category of their own. Cloudflare adds multimodal input, open weights and a fine-tuning service, though Clef costs several times more per token than Jev.
Changelog
- 2026-10-02: created
Models
- Clef-flash Cloudflare · current
- Clef Cloudflare · current
Related posts (1)
- Cloudflare's Workers AI team introduces Clef, its first trained models original ↗ michelle @michellechen · x · 2026-10-01
First-hand launch post from Cloudflare's Workers AI team for the Clef decision models (~171k views), higher reach than the corporate post.
Related events
Sources (6)
- officialCloudflare blog: Clef decision models
- docsCloudflare Workers AI docs: clef
- codeHugging Face: Cloudflare/clef
- officialmichelle (Cloudflare Workers AI) on X: introducing clef (~171k views)
- pressMarkTechPost: Cloudflare releases Clef and Clef-flash
- pressThe Register: Cloudflare tries to outplay Jev with open-weight Clef models
id: 2026-10-01-cloudflare-clef-decision-models · updated 2026-10-02 · open in the interactive timeline