Models & inference
Clef
Cloudflare's open-weight decision models that return typed answers with probabilities, hosted on Workers AI.
What Clef does
Clef is a pair of decision models from Cloudflare, Clef and Clef-flash, announced on the Cloudflare blog on October 1, 2026. A decision model does not write free text. You give it an input state and a set of typed questions, and it returns a probability for each allowed answer, which your code can use to route a ticket, trigger an escalation or hand the case to a person.
Cloudflare's docs list Clef at 27B parameters and Clef-flash at 9B. Both read the state as text, JSON, images or video, have a 65,536-token context window, and take up to 64 questions per request in three types: yes/no, choice and score. They run on Workers AI as @cf/cloudflare/clef and @cf/cloudflare/clef-flash, and Cloudflare says they are compatible with the Jev API from Typesafe AI. The weights are on Hugging Face under Apache 2.0, and the model card names Qwen3.8-27B as the base for Clef.
Cloudflare also announced a reinforcement learning fine-tuning service. It starts as hands-on work with Cloudflare's forward-deployed engineering team, with a self-serve platform planned later. The announcement says Cloudflare does not read, store or train on Clef requests and responses unless a customer uses fine-tuning.
Clef suits developers who need fast, bounded classification steps inside agent workflows. The benchmark tables in the announcement are Cloudflare's own.
- Decide whether a support message is urgent and which team should handle it.
- Classify a website domain into categories, as Cloudflare's Threat Intelligence team does.
- Classify images or video frames against a fixed set of labels.
- Add a decision step to an agent workflow that can defer to a human when confidence is low.
Who should use it
- Developers on Cloudflare Workers who need classification inside agent workflows.
- Teams that want open weights they can run themselves.
Who should look elsewhere
- Chat or open-ended text generation, since the models output only probabilities.
- Teams that want self-serve fine-tuning now; the service starts as engineer-led work.
Strengths
- Accepts images and video as well as text and JSON.
- 65,536-token context window on both models.
- Apache 2.0 weights on Hugging Face, plus a hosted API on Workers AI.
- Hosted pricing is per million input tokens, with a free daily Workers AI allocation.
Limitations
- Answers are limited to the options you define; the models do not generate text.
- Benchmark results come from Cloudflare's own announcement.
- Fine-tuning is not self-serve yet.
Sources
- Introducing Clef: our open-source decision models, and new RL fine-tuning platform | Cloudflare BlogOfficial sitePrimaryblog.cloudflare.com
- Posts tagged "Plans" — Cloudflare BlogPricing pagePrimaryblog.cloudflare.com
- Posts tagged "Developer Documentation" — Cloudflare BlogOfficial docsPrimaryblog.cloudflare.com
- developers.cloudflare.com/workers-ai/models/clefOfficial docsPrimary
- developers.cloudflare.com/workers-ai/models/clef-flashOfficial docsPrimary
- developers.cloudflare.com/workers-ai/platform/pricingOfficial docsPrimary
- huggingface.co/Cloudflare/clef/blob/main/LICENSERepositoryPrimary
- huggingface.co/Cloudflare/clefRepositoryPrimary
Clef alternatives
PocketWebTools for Mac
Models & inference
Mac app that runs open AI models up to 35B parameters offline for chat, transcription, OCR, upscaling and video clipping.
PaidVerifiedej
Models & inference
An 11 MB open-weight model from HITL Systems that answers routing, yes/no and priority questions about text on CPU.
Open sourceVerifiedLloyal
Models & inference
A TypeScript framework for AI apps that run an open-weight model inside the app, offline and without an API key.
Open sourceFree tierVerifiedReplicate
Models & inference
A cloud API for running, fine-tuning and deploying open-source and proprietary AI models, billed by usage.
PaidAPIVerifiedLM Studio
Models & inference
Desktop apps for running open LLMs locally, with a local API server and Bionic, an agent for documents, code and automation.
FreemiumFree tierAPIVerifiedOllama
Models & inference
An open-source tool and API for running open models on your own machine or on Ollama's cloud.
Open sourceFree tierAPIVerified
Related open-source projects
OpenMontage is a Python toolkit that turns an AI coding assistant such as Claude Code, Cursor, Copilot, Windsurf or Codex into a video production setup. You describe the video, and the assistant moves through fixed stages: research, proposal, script, scene plan, assets, edit and compose. No separate orchestrator program runs the show. The assistant reads YAML pipeline manifests and a Markdown director skill for each stage, then calls the Python tools itself. The repository has pipelines for animated explainers, cinematic trailers, documentary montages, talking-head edits, screen demos, podcast clips, and localization with dubbing, among others. It connects to many providers, including Kling, Veo, Runway, FLUX, ElevenLabs and Suno, and to local options: Piper for speech, WAN or Hunyuan video models on a GPU, and Remotion, HyperFrames and FFmpeg for rendering. With no API keys at all it can still narrate with Piper, pull footage from Archive.org, NASA and Wikimedia Commons, and render with Remotion. Creative stages wait for your approval, and a local board called Backlot shows each stage, the generated assets and the money spent. Budget controls estimate costs before generation, with a default total cap of $10 and a confirmation step for any single action above $0.50.
- Stars
- 68K
- Language
- Python
- License
- AGPL-3.0
- Momentum
- +8.0% 7dEstablished
qwen-audio-agent is a realtime voice runtime that sits in front of a coding or task agent. You talk to a voice model, longer jobs go to a backend agent while the conversation continues, and when a task finishes its result comes back into the same conversation. The voice side is swappable. Cloud options listed in the README include Qwen Audio 3.0 Realtime, GPT-Live / OpenAI Realtime, Google Gemini Live and Doubao Seeduplex 3.0 Realtime, each needing that provider's API key. Hugging Face Speech-to-Speech and MiniCPM-o 4.5 can run locally. Backend agents connect mostly over ACP: the list covers Qwen Code, OpenCode, OpenClaw and Kimi Code, with Codex and Claude Code joining through an external ACP adapter. You can use it from a WebUI, a terminal TUI or a desktop floating orb on macOS, Windows and Linux. The README also lists parallel background tasks you can query or cancel, cross-session memory, and scheduled reminders. The package is published on npm under Apache-2.0, a technical report is on arXiv, and the repository sits in the QwenAudio GitHub organization.
- Stars
- 2.9K
- Language
- JavaScript
- License
- Apache-2.0
- Momentum
- Emerging
iFixAi is an open-source Python tool that audits AI agents and models by running a set of test inspections and returning an A to F grade with a written report. The free engine has 60 inspections. The grade comes from five core pillars: fabrication, manipulation, deception, unpredictability and opacity. Another 28 inspections from twenty premium categories are included and reported on their own, outside the grade. A grade counts as citable only when a judge model from a different vendor scores the run, so a full audit needs two API keys. You can test a bare model through providers such as Anthropic, OpenAI, Gemini, Azure, Bedrock and OpenRouter, or point the tool at your own agent's OpenAI-compatible HTTP endpoint. There are three ways to run it: a setup wizard, a CLI with explicit flags for CI, or a plugin for Claude Code and Codex, with a scaffolded /ifixai-skill command for other agents. Reports land as JSON and Markdown in ./ifixai-results/. The README estimates judge costs for a full run with a Sonnet judge at roughly $12 to $18, billed by the provider. The company also sells iFixAi Pro, a paid audit service with up to 400 inspections. Telemetry is on by default outside CI and can be switched off.
- Stars
- 24K
- Language
- Python
- License
- Apache-2.0
- Momentum
- Established
AI Engineering from Scratch is a free, MIT-licensed course on GitHub that teaches how AI systems work by having you write each piece yourself, starting with linear algebra and ending with multi-agent systems and production infrastructure. The material is split into 20 phases and 523 lessons, with code in Python, TypeScript, Rust and Julia. Each lesson folder holds a written explanation, runnable code and an outputs directory. The pattern is to build an algorithm such as backpropagation, a tokenizer, attention or an agent loop from plain math first, and only then run the same idea through a library like PyTorch or scikit-learn. Lessons end with something you keep: a prompt, a SKILL.md file, an agent or an MCP server. The repository holds 396 skills and 99 prompts produced this way, plus a script that copies them into an agent's skills folder. There are three ways to follow it. You can read lessons on the website with no setup, clone the repository and run the code locally, or install the course's tutor skills into a coding agent such as Claude Code or Codex. Those skills run a placement quiz, write a study plan to LEARNING.md and teach one lesson per session. Separate tracks prepare for the four Claude certification exams and the MCP Associate exam; the README says these are independent and not affiliated with the exam bodies.
- Stars
- 68K
- Language
- Python
- License
- MIT
- Momentum
- +8.2% 7dEstablished
Related guides
- News
Cloudflare releases Clef decision models and a managed RL fine-tuning service
Cloudflare released Clef and Clef-flash, decision models that return typed answers with probabilities, on Workers AI and Hugging Face. Prices, specs and gaps.
4 min read