Clef decision models ship Apache 2.0 weights, a 38.8 ms flash tier and RL fine-tuning

A decision model takes input state and a set of typed questions and returns bounded answers with probabilities. Code can use those answers to route a ticket, escalate, or defer to a human. The post names two concrete differences between Clef and Jev: Clef has a vision encoder, so it can classify images, where Jev only does text today, and it has a 64k context window against Jev's 32k. Cloudflare says it does not read, store, or train on hosted requests or responses unless a customer opts into fine-tuning.

Cloudflare shortlisted ten benchmarks it considers important under the Jev Decision Index. It scored Clef, Clef-flash, Jev, DiffusionGemma Jev, Kev 9B and Laya on them. These are Cloudflare's own runs, also published on its live benchmark demo site, not an independent evaluation.

Cloudflare also ran Typesafe's own workflow eval suite. Clef beat Jev on invoice processing (64.7 vs 61.8) and security incidents (62.9 vs 61.7, where Clef-flash tied Jev at 61.7). Clef-flash edged Jev on customer service (77 vs 76.0), and Jev kept agent trace observability (71.6 vs Clef-flash 69.8 and Clef 68.5). Every one of those four gaps between Jev and the best Clef variant is 2.9 points or less. A team's own labelled sample will tell it more than the headline that Clef wins 3 of 4.

Clef runs the Qwen backbone as a prefill-only pass, then scores the valid schema choices in parallel. The decision step is non-autoregressive, so no intermediate text is generated. A two-stage attention routing process lets each option pull in prompt context and lets fields cross-attend to each other before schema-bound scoring. Across 43 eval benchmarks, Cloudflare reports median latency of 209.3 ms for Clef, 38.8 ms for Clef-flash, 524.1 ms for Jev, 84.4 ms for DiffusionGemma Jev, 51.4 ms for Kev-9B and 5.8 ms for Laya. By that table, Clef-flash beats every model except Laya on median latency, but the larger Clef is slower than Kev-9B and DiffusionGemma Jev.

Clef's p95 of 238.6 ms sits close to its 209.3 ms median, while Clef-flash's p95 rises to 122.4 ms from a 38.8 ms median. Even so, Clef-flash has the lowest p95 in the table, against 536.0 ms for Jev and 222.5 ms for Laya. In a domain classification workflow with Browser Run, Cloudflare's Threat Intelligence team reports Clef took 2.2 s to fetch, render and classify a site. gpt-oss-120b took 4.7 s in the same workflow and returned only two classifications. Cloudflare says Laya is faster but trades away quality on its benchmarks.

The hosted model sits at @cf/cloudflare/clef on Workers AI. A request carries a state string plus named questions, typed in the post's example as choice, score, and a field type written as noul for a yes-or-no question. Cloudflare says Clef produces strictly typed outputs and is fully API-compatible with Jev, so the swap should be easy. The post does not show the endpoint or model id for Clef-flash, and the noul type name should be checked against the developer documentation before copying the example.

Cloudflare positions Clef as the precision model and Clef-flash for latency-critical decisions. It says GPUs at the edge cut network latency, so the call can sit in an agent's hot path next to a Workers AI LLM that takes the action. The blog post does not state pricing. One Hacker News commenter put Clef at $0.24 per million input tokens, roughly 6x Jev, and Clef-flash at $0.09. Treat those figures as unverified until they are checked against Cloudflare's pricing page.

Cloudflare's stated case for fine-tuning is specific workloads with a long history of labelled decisions. Its examples are Trust and Safety submissions, Support triage, and good-bot versus bad-bot calls in its Bot products, drawing on more than 15 years of network data. Cloudflare says such a model can be more accurate and faster than generic Clef, but it publishes no numbers for a fine-tuned model. The post is explicit about the trade: a fine-tuned model may give up general-purpose performance for higher accuracy in one domain. If inputs and categories shift often, the hosted general model fits better, since decision models are meant to handle new categories without retraining.

The base Clef recipe jointly trains a routing head and rank-256 low-rank adapters over the frozen backbone. It uses label-smoothed cross-entropy plus a Brier loss for calibration, on synthetic data that permutes field orders, prompts and schema structures. Reinforcement Learning for Calibrated Decisions (RLCD) serves as a secondary optimization target: it gives partial credit for adjacent ordinal choices, rewards fully precise record outputs, and applies a reference penalty against distribution shift. The customer fine-tuning pipeline is assembled from Cloudflare platform primitives, one of them new:

Cloudflare describes several of these pieces as work in progress. Today, Clef RL fine-tuning is a hands-on engagement with the FDE team, and the self-serve platform comes later, with no date given. Cloudflare also invites existing customers of these products who have specific use cases to act as design partners.

The top-ranked thread asked how so many Jev alternatives appeared within weeks. Replies largely agreed that the technique is not new. One commenter said any pretrained LLM can be adapted to score a structured set of options, and several said Jev's contribution was the API and product concept. Cloudflare's own account fits that: its first prototype adapted DiffusionGemma by exposing logprobs, before Clef moved to a Qwen backbone.

The post omits several things an engineer would need. It gives no calibration numbers despite training explicitly for calibration, no before-and-after results from any fine-tuned Clef, and no hardware requirements for running the 27B-backbone model locally. It also gives no date for the self-serve RL platform and does not state the hardware or network conditions behind its latency figures. All accuracy and latency figures come from Cloudflare's own runs on benchmarks it chose. Before swapping out Jev, run both against a labelled sample of your own decisions, watch the When2Call and CLINC150+OOS gaps if your workload resembles them, and confirm pricing directly.