Cloudflare open-sources the Clef models, aiming straight at Jev

Share
Cloudflare open-sources the Clef models, aiming straight at Jev

Cloudflare picked October 1 to wade into the fastest-moving model category of the fall: bounded, machine-consumable decisions rather than prose — and it brought open weights and a price list.

Cloudflare released Clef and Clef-flash, a pair of open-weight "decision models" built to beat TypeSafe's Jev at its own game. The category is Jev's invention — take in unstructured state, return typed, probabilistic decisions that software can act on, with no free-text sampling to hallucinate. Cloudflare's pitch is a drop-in: both models are Jev-API compatible, live on Workers AI, and ship under Apache 2.0, with Clef built on Qwen3.8-27B and Clef-flash on Qwen3.5-9B. On latency, Cloudflare claims a 209-millisecond median for Clef and 38.8 milliseconds for the flash variant, against 524 milliseconds for Jev — numbers that matter enormously for the workflow automation this category targets. Cloudflare also opened a fine-tuning service so buyers can adapt the models to their own decision taxonomies. Two more spec edges are worth noting: 64k context windows where Jev caps at 32k, and a vision encoder that lets Clef classify images and video in the same decision flow — something Jev, text-only, cannot do. Running either model yourself is not cheap, though: 41 GB of VRAM for Clef-flash, 85 GB for Clef.

The honest caveat: every score is Cloudflare's own, and none of it has been reproduced for the official Decision Index rankings. The split is instructive anyway — Clef wins the intent-classification tests handily (94.20 vs 79.74 on BANK77), while Jev still crushes it on knowledge-heavy reasoning (GPQA Diamond, 78.3 vs 48.0). Pricing cuts the other way: Clef charges $0.24 per million input tokens, nearly six times Jev's $0.042. And "open source" has an asterisk — the weights are public, but Cloudflare confirmed the training datasets are not. Our take: the interesting move isn't the benchmark tennis, it's that a major infrastructure vendor is treating decision models as a product line worth funding — the second open challenger to arrive since Jev's launch, after we covered Kev's open Jev alternative now trains on Qwen3.5 for about $95. A category gets real when the clones arrive; the ones who will feel it first are the vendors who priced on scarcity.

What to watch: whether the official Decision Index reproduces Cloudflare's numbers, and whether Jev responds on price.

Does open weight-plus-API-compatibility beat a head start in a category this young? Tell us in the comments.

Read more

OpenAI drops 372 math results, nearly all from one prompt

OpenAI drops 372 math results, nearly all from one prompt

A dense news day across three fronts: OpenAI published a mountain of mathematics from an unreleased model, SpaceX reportedly wants the debt market to fund its Nvidia orders, and Anthropic restructured its cyber program around three tiers of trust. OpenAI published 372 mathematical result families from an internal frontier model — and says nearly all of them came from a single prompt handed to a single AI agent. The release comprises 722 manuscripts touching hundreds of open questions, plus Lea

Underdog launches a private on-device AI assistant, backed by a16z

Underdog launches a private on-device AI assistant, backed by a16z

The privacy split in consumer AI got a new entrant tonight, GitHub turned code-review benchmarks into a vendor-bias argument, and OpenAI opened its oddest API to everyone. Underdog launched in invite-only beta: an on-device AI assistant that keeps your data on your machine and never charges you a subscription. Self-taught coder and Thiel Fellow Sigil Wen — who moved to Silicon Valley at 17 and lived in an AI hacker house with Andrej Karpathy — built his own inference engine, Husky, to run a 2

Today in AI — October 6, 2026

Today in AI — October 6, 2026

A day of second-order moves: the labs are buying task data from software vendors instead of scraping the web, Waymo is borrowing to fund the robot world, and the biggest bank in the US just put a number on what Anthropic's latest model cost it in risk. Models & Research * OpenAI is training GPT-6 Astra on Ironclad's real contracting work. Ironclad staff helped turn 11 tasks across legal, commercial and procurement work — setting up NDAs, approval workflows, clauses that change by jurisdict

Meta, Walmart and Stripe publish the Personal Agent Protocol

Meta, Walmart and Stripe publish the Personal Agent Protocol

The agent economy is writing its rulebook tonight: one open standard for AI bots at the checkout, a cheaper image model from Google, and Anthropic turning bug-hunting into a tiered product. Meta, Walmart, Stripe and Sierra are publishing an open "personal agent protocol" — a standard that defines how personal AI agents interact with businesses online. The group behind it reads like a cross-section of agentic commerce: Meta, Sierra, Walmart, Stripe, Shopify, Genesys, Rocket, NiCE, Decagon and I