nameDataxzuzf9TH1OA
titleData
Clef: Open-source decision models, and new RL fine-tuning platform embedTexthttps://www.youtube.com/embed/xzuzf9TH1OA
youtube_urlDatahttps://www.youtube.com/watch?v=xzuzf9TH1OA
descriptionLong TextThirty-eight point eight milliseconds — a median, not a best case. Cloudflare published it for Clef-flash, one of two open-source decision models it announced on October 1, 2026. In the same latency table, Jev from Typesafe AI sits at 524.1 ms. Roughly thirteen times.
A decision model isn't a chatbot. A language model answers in free-form text; Clef returns a typed, probability-scored answer against a schema you define. Cloudflare's own example: one domain scored 95% chance fashion website, 85% ecommerce, under 1% phishing. Three fields, three probabilities, one HTTP POST. Clef and Clef-flash run on Workers AI, are open-sourced on Hugging Face under Apache 2.0, and are described as fully Jev-API compatible.
The interesting part is where the milliseconds come from. Clef freezes Qwen3.8-27B (Qwen3.5-9B for Clef-flash) and trains a routing head plus rank-256 low-rank adapters on top. At inference it runs a prefill-only pass, then scores the valid schema choices in parallel: "The decision step is non-autoregressive, so there's no intermediate text to generate token by token." No decode phase, no token queue. A Brier loss in post-training is what should make a 95% actually mean 95 times in 100.
The scoreboard, from 43 eval benchmarks: BFCL case exact 98.47 (Clef) and 98.76 (Clef-flash) against Jev's 95.75; API-Bank accuracy 91.93 and 93.11 against 88.19; BANKING77 macro-F1 94.20 against 79.74. The detail I respect most is that Cloudflare published the losses too — Jev leads When2Call 80.97 to 72.37 and BRIGHT nDCG@10 47.52 to 45.91.
What isn't published yet, said plainly:
- No pricing — not per request, per token or per decision.
- No Hugging Face repo names, parameter counts or local hardware requirements.
- No GA-versus-beta answer, and no word on whether vision is live in the hosted API.
- No paper and no ablations for the two-stage attention routing architecture.
⚠️ This is a spec-sheet reading of one published announcement — nothing was installed, benchmarked or reproduced here, and every figure quoted is Cloudflare's own.
⏱️ Chapters:
00:00 Thirty-Eight Point Eight Milliseconds
00:32 Two Models And A Platform
01:03 What Is A Decision Model?
01:28 Number One: The Latency Table
02:02 Read The P95, Not Median
02:40 A Decision You Can Loop
02:58 Number Two: Frozen Qwen Backbones
03:26 Clef Skips The Decode Phase
03:54 A Queue Versus A Grid
04:24 Number Three: Brier Loss Calibration
04:56 RLCD And Partial Credit
05:27 Number Four: BFCL And API-Bank
06:04 BANKING77 And The Widest Gap
06:59 The Benchmarks Cloudflare Does Not Lead
07:55 Number Five: 64k Context, Vision
08:18 The Call Is One POST
08:53 Number Six: The Browser Run Test
09:24 2.2 Versus 4.7 Seconds, Precisely
09:51 Number Seven: The Fine-Tuning Loop
10:24 Cloudflare's Own Three Use Cases
10:52 The vLLM Credit Worth Reading
11:15 Two Things Are Real Today
11:55 What Is Not Published Yet
12:44 Who Clef Is Actually For
13:23 Where The Name Clef Comes From
13:36 Now The Numbers Are Falsifiable
🔗 Sources:
- Announcement: blog.cloudflare.com/clef-decision-models/
- Workers AI docs: developers.cloudflare.com/workers-ai
- Licence: apache.org/licenses/LICENSE-2.0
#Cloudflare #Clef #WorkersAI #DecisionModels #SignalCoders
👇 Subscribe — a new plain-English breakdown of AI engineering, open source tools, and local LLMs every single day.
YouTube: www.youtube.com/@Signalcoders
Twitter/X: @SignalCoders
Our Playlists:
https://www.youtube.com/playlist?list=PLL70bDS-hH7Y
https://www.youtube.com/playlist?list=PLTnqXOgLPogE
https://www.youtube.com/playlist?list=PLZVB6navai50
https://www.youtube.com/playlist?list=PLEj9VzjWmmho
⚠️ Disclaimer: Educational purposes only. Signal Coders is not affiliated with or sponsored by Cloudflare, Typesafe AI, Hugging Face, Qwen, OpenAI, Replicate, the vLLM project or any person or project mentioned. This is a reading of a single public blog post by Michelle Chen, Alex Reneau and Kevin Flansburg as it stood on the recording date, quoted verbatim where quoted — the models were NOT installed, run, benchmarked or tested here, and NO figure was reproduced. All latency, benchmark, context-window and workflow numbers (including the 43 evals, the BFCL, ToolRet, API-Bank, BANKING77, CLINC150, When2Call, BRIGHT and PhishNChips scores, the Jev Decision Index placement and the 2.2 vs 4.7 second Browser Run comparison) are Cloudflare's own, measured on Cloudflare's infrastructure under conditions it did not fully specify; competitor serving conditions are unstated. The data-handling statement quoted is Cloudflare's stated policy, not an external audit. Pricing, availability, repo names and parameter counts were not published at recording time — verify current docs and licence terms before building on any of this. Readings of the architecture and of what is missing are my opinions, not measured findings. Nothing here is professional, security or financial advice; consult qualified professionals for systems that matter.
published_dateDate2026-10-02 20:31:50
is_publishedCheck0