Cloudflare Ships Open-Weight Clef Models to Unseat Jev

Cloudflare released Clef and Clef-flash, Apache 2.0 decision models that beat TypeSafe's Jev on most benchmarks while running up to 13 times faster.

@sharedot

Cloudflare Ships Open-Weight Clef Models to Unseat Jev

By @sharedot · · 8 pages

  • AI Frontier
  • Decision Models
  • Open Weights
  • Cloudflare Clef
  • Jev

Cloudflare released Clef and Clef-flash, Apache 2.0 decision models that beat TypeSafe's Jev on most benchmarks while running up to 13 times faster.

What happened: a decision model, not a chatbot

Cloudflare's Workers AI team has released its first two trained models, Clef (27B) and Clef-flash (9B), and both are decision models rather than chatbots.

Each reads an input state and a schema of typed questions, then returns a probability for every allowed answer with no free-form text — supporting yes/no (noul), named-choice, and rubric-scored question types.

Both weights are on Hugging Face under Apache 2.0, run today on Workers AI, and are fully compatible with TypeSafe AI's Jev System One API, so teams can switch by changing the endpoint and model name.

MarkTechPost reports the models are post-trained from Qwen3.8-27B and Qwen3.5-9B respectively, keeping the backbone's vision encoder.

Why it's surprising: skipping text generation entirely

The surprising technical move is that Clef never generates output tokens one at a time the way an LLM does.

MarkTechPost describes a two-stage inference: the Qwen backbone runs a single prefill-only pass over the state and questions, then a small joint schema head reads the final hidden states, routes evidence to each question, lets fields cross-attend, and scores all options jointly before a per-question softmax turns logits into probabilities.

Pasquale Pillitteri's coverage explains the training recipe: backbones were frozen while the routing head was optimized with rank-256 low-rank adapters, pairing label-smoothed cross-entropy with a Brier loss for calibration, plus Cloudflare's RLCD technique that gives partial credit for adjacent ordinal choices.

The result is schema-bound scoring at a fraction of typical inference cost.

The evidence: benchmark wins — with caveats

On Cloudflare's 10-benchmark shortlist from the Decision Index 0.2.1 suite, a Clef model scored highest on 7, and the numbers can be lopsided: BANKING77 macro-F1 came in at 94.20 for Clef versus 79.74 for Jev, and Clef-flash hit 97.73 on home-appliance case-exact against Jev's 52.27.

Jev keeps clear leads on knowledge-heavy tests — GPQA Diamond 78.3 versus 48.0, MMLU-Pro 82.7 versus 65.9, and BBH 92.9 versus 73.7 — and Pillitteri adds that Jev stays ahead on When2Call and BRIGHT. Latency is where Clef-flash shocks: 38.8 ms median versus Jev's 524.1 ms.

In Cloudflare's own threat-intelligence workflow, Clef classified a domain in 2.2 seconds against 4.7 seconds for gpt-oss-120b.

All figures are vendor-reported with no independent replication yet, a caveat The Register flagged via Streamline's reporting, which also notes hosted pricing of 24 cents per million tokens for Clef versus 4.2 cents for Jev.

A crowded field forms in two weeks

Startup Fortune reports OpenAI introduced a Decisions API at Dev Day on September 30, built on its Luna model, and Amazon shipped Strands Decider 2B on October 1 — a 2B-parameter Qwen3.5-based model from Strands Labs that Cloudflare's release leaves as the only one of the three Jev challengers giving away full weights, with OpenAI's API hosted and Amazon's model small.

Startup Fortune also reports that Jev's cost profile drove the rush: a hackathon demo reportedly monitored an AI agent's actions for $2.94 versus $372 through a frontier LLM.

Fastino Labs separately released GLiDE, which it bills as the first 'thinking' decision model, claiming 64.81 on Decision Index 0.2.1 versus Jev's 57.91.

The stakes: lock-in versus self-hosting

Startup Fortune frames the real contest as infrastructure, not model quality: a team building an agent pipeline against OpenAI's closed Decisions API bets its workflow on one company's uptime and pricing, while a team that pulls Clef's weights from Hugging Face owns its copy forever.

That matters because decision-making is becoming its own layer of the stack — Fastino Labs CEO George Hurn-Maloney compares it to retrieval, saying you shouldn't need a frontier model every time an agent picks its next step.

There are operational costs either way: Streamline reports self-hosting a 64,000-token context at single concurrency needs about 85GB of GPU memory for Clef and 41GB for Clef-flash, tested on a single H200 with BF16 weights.

Clef also uniquely accepts images alongside text, something neither Jev nor Amazon's Decider claims.

What comes next: fine-tuning and vision

Cloudflare is pairing the release with a reinforcement learning fine-tuning service for tuning Clef on private data, built from AI Gateway, Workers AI, Containers and a new Trainer component that redistributes tuned models onto Workers AI via Bring Your Own Model.

Pillitteri reports it launches in white-glove mode with Cloudflare's forward-deployed engineers before opening to self-serve, and that Cloudflare says it does not read, store or train on customer request and response data unless the customer opts into the fine-tuning product.

Meanwhile, beyond decision models, the open-weight wave is spreading to robotics: Robotics & Automation News reports Runway is testing Praxis-1, its first open-weight 'world action model' trained on web video rather than teleoperated robot data, with partners Noble Machines, Standard Bots and Ultra Robotics ahead of a public release in coming months.

Sources

  1. marktechpost.com › Cloudflare Releases Clef and Clef-flash: Open-Weight Decision Models That Return Typed Probabilities Instead of Text
  2. startupfortune.com › Cloudflare Open Sources Clef to Challenge OpenAI and Amazon on AI Agent Decisions
  3. pasqualepillitteri.it › Cloudflare launches Clef, the model challenging Jev on AI decisions
  4. streamlinefeed.co.ke › Cloudflare Clef Brings Open Weights to AI Decisions
  5. prnewswire.com › Fastino Labs Releases GLiDE, the First Thinking Decision Model
  6. roboticsandautomationnews.com › Runway moves into robotics with open-weight Praxis-1 AI model

More on AI Frontier