Cloudflare ships Clef, an open decision model with pictures

Meta description (draft): Cloudflare open-sourced Clef: a 27B scoring head on a frozen Qwen3.8-27B that scores answer sets, takes images, and returns probabilities in 209 ms. —

Cloudflare open-sourced two decision models on October 1: Clef and Clef-flash, Apache-2.0 on Hugging Face, with weights on the Hugging Face repos and code paths documented in Workers AI. Clef is a scoring head and rank-256 LoRA on a frozen Qwen3.8-27B backbone; Clef-flash rides a frozen Qwen3.5-9B. Both return probabilities for fixed answer sets instead of generated text, and both accept images, which no other decision model in this lane does today. Independent aggregators noticed within a day; the table that circulated on X (Mia’s screenshot, 534 likes) matches the blog’s own benchmark table number for number across all ten rows, so the figures below carry the blog’s data with the caveat attached.

On the vendor’s own Decision Index leaderboard (marked self-reported, and The Register notes independent verification is still pending), Clef currently tops the class at 61.21 against Jev’s 57.91. Latency is the stronger claim for daily use: 209 ms median for the 27B model and 38.8 ms for the flash, against 524 ms for Jev on the same harness. Cloudflare’s own developer account frames the comparison more bluntly: “Like Jev? You’ll love Clef and Clef-flash, which are smarter, faster, and fully Jev-API compatible” (@CloudflareDev), and the vendors’ numbers behind it are the two lines above: latency 2.5x better against the full model and 13.5x against the flash, index accuracy 3.3 points, with the outlier rows wider (Home appliances 82.95 versus 52.27).

*Median decision latency on the same harness: the flash variant answers a multiple-choice prompt in 38.8 ms while the full model takes 209 ms and Jev 524 ms. All three are vendor-measured.*

How it works

The mechanics follow Kev’s approach with a multimodal twist. The frozen Qwen backbone does a prefill-only pass; nothing autoregressive happens. A separate routing head then scores every valid option of every schema-defined question in parallel, up to 64 questions per request. The routing head is a two-stage attention design: option-specific evidence extraction followed by joint cross-field attention, with a lexical prior that keeps intent anchored when you swap option phrasings.

Clef-flash keeps the vision encoder from its backbone: PNG, JPEG, WebP, up to 4 images per request, each capped at 4 MiB and 16 MP, with 8 MiB total decoded per request. Training used label-smoothed cross-entropy for the schema outputs plus Brier loss for probability calibration, and a secondary objective the Cloudflare team calls RLCD, Reinforcement Learning for Calibrated Decisions: partial credit for adjacent ordinal picks, a reward shape for fully precise outputs, and a reference-model penalty to stop distribution drift.

Self-hosting limits are concrete: the Cloudflare PM told The Register the flash needs at least 41 GB VRAM and the full model at least 85 GB, at single concurrency and 64k context. That puts the full model on the workstation tier (96 GB boxes) and the flash on enthusiast hardware; the hosted API is $0.24 per million input tokens for Clef and $0.09 for Clef-flash.

Strands Decider 2B: the AWS entry

One day earlier, AWS’s Strands Labs shipped Strands Decider 2B, and the origin story shows where this category is going: Marc Brooker saw TypeSafe’s Jev and built his own, then Amazon shipped it. Strands Decider 2B is a rank-16 LoRA and pointer head (about a million parameters of fresh structure) on a frozen Qwen3.5-2B base, with the training scripts and datasets published in the repo. On JevBench it ranks 3rd of 33 models in its size class and first of 30 excluding the just-over-2B entrants, with a 72.29 percent public-set accuracy in the model card’s own eval table.

*Size on one axis, latency on the other: Strands Decider 2B plays in a different part of the chart than Clef, trading score for hardware reach.*

The week’s shape

One week after Jev launched, the lane has a TypeSafe frontier, an open Apache-2.0 challenger from one CDN with images and a 64k window, and a 2B entry from a hyperscaler that runs on a Pi-class box. Ollama meanwhile ships the Jev API locally in 0.35 with Nimble and Tev1 pulls. Every one of these is a bet that most agent decisions are classification, not generation: model routing, tool selection, guardrails, triage, and evaluations all reduce to structured choices with confidence scores attached. The local case is not price anymore; it is the sub-100-ms loop and owning the decision boundary inside your own harness.

Open weights win this round on fit math: Clef runs a 55 GB checkpoint for the price of a 96 GB box, Clef-flash runs where Qwen3.5-9B runs, Strands Decider 2B runs on a CPU. If the Decision Index scores hold under independent verification, the classification workloads that quietly make up most of an agent’s bill stop being a frontier-model cost and start being a local line item.

Update, October 4: adoption counts and the one-card correction

Four days in, the adoption is real: Clef sits at 1,073 likes with 4,214 downloads, Clef-flash at 386 likes and 6,372 downloads, and the community quant scene did to Strands Decider 2B what it does to everything open now: mradermacher’s GGUF set at 3,770 downloads, Mapika’s at 2,049, cosetoenor’s at 1,850, plus vision and DreamBlooms variants, all inside 48 hours of release. The category has its day-one ports.

A 4-bit Clef is about 14GB of weights, but the KV cache is the budget line, because the 64-layer Qwen backbone runs GQA at 4 heads x 256 dims, which is 262MB of cache per 1,000 tokens at fp16. At the full 64k context that is 16.4GB on top of the weights: 30GB total, past a 24GB card. The honest one-card statement: a 24GB 4090-class card runs 4-bit Clef at short-to-mid context (up to roughly 30k tokens), a 48GB pair of 3090s covers the full 64k, and the flash model’s smaller cache is part of why it is the practical single-card pick. Check the context window you actually serve before believing any one-card claim, including ours.

The land grab, one table

Five weeks of decision-model announcements: TypeSafe’s Jev opens the category mid-September; Ollama ships the Jev API locally September 28 with Nimble and Tev1; Cloudflare ships the open multimodal scorer September 30; Amazon’s Strands ships the open 2B September 30; OpenAI showed a closed Jev-style model at DevDay and AWS released Strands a day later per TechCrunch. Every major inference platform and one hyperscaler now runs a bet that most agent decisions are classification. The open ones publish weights; the closed ones publish pricing; your box runs either the same day.

Sources: Cloudflare blog - Workers AI clef docs - Cloudflare/clef on HF - Cloudflare/clef-flash on HF - Strands Decider 2B blog - StrandsAgents Decider on HF - The Register on the Clef launch - Cloudflare Developers’ Clef quote-tweet

The reckoning: Red Hat’s benchmark says the category isn’t magic

The strongest public test landed October 2: Red Hat’s AI Safety team benchmarked Jev-style decision models against traditional guardrails and found no crown. Across two guardrail tasks, 89.01 percent accuracy from a 200M-parameter CPU classifier, 89.31 from a 35B LLM judge, and Jev itself at 86.35 with a 348ms median that includes a transatlantic round-trip. Laya, the 421M-parameter Jev-style model, hit 85.44 accuracy at 119ms median; on content safety it collapsed to 57.87 accuracy, the weakest of every entry except BART. Red Hat’s own conclusion: pre-trained predictive models remain the top pick where labeled data exists, and decision models are a capable zero-shot alternative, not a faster-cheaper-better replacement. Their closing note is plain and worth carrying: the Jev wave refocuses attention on lightweight, task-specific inference, and that pragmatism is a win regardless of which paradigm wins a benchmark. What the benchmark does not cover matters as much as what it does: agentic steering, form-parsing, and routing on state objects - the workloads Clef actually ships for - are absent; guardrails on labeled risk corpora are where traditional classifiers are strongest.

Jerry Gamblin’s independent Clef test points the same direction: Clef-flash read all 27,489 CVE descriptions from the last 60 days on his laptop at under a second each, no API key, no bill - and the 4.2 percent of non-kernel CVEs (989 of 23,736) that never state a security impact is a real signal the description-quality problem exists. The two-step clean-up (Flash flags, the 27B re-reads and overturned 54.8 percent of its flags) against Gamblin’s own 129 hand labels is exactly the shadow-then-authority playbook production ML has always needed: the model is cheap, the trust still has to be earned.

The first viral consumer artifact arrived a day later: Carson Jones tagged every NHL goal this season - 200-plus - with Clef over one weekend, extracting shot type, strength state, celebrations, scrums, and per-player confidence scores from frames, for under two dollars of Workers AI spend. No fine-tuning, no harness: a schema of fixed-option questions per goal clip, and every tag came back as a scored option from his option list. It is the clearest demonstration yet of the category’s promise: one fan, one weekend, a video-understanding pipeline that would have been a CV-team project in 2024.