Clef: Score structured decisions with Clef
Added · Updated
Made by Cloudflare
View profile on Hugging Face (opens in a new tab)Ask yes-or-no, multiple-choice or rating questions about a piece of text, JSON, an image or a video, and get a probability for every allowed answer.
It is for developers who build routing, triage or review steps into their apps. Clef is a 27-billion-parameter multimodal model (it reads images and text together) from Cloudflare, trained further from Qwen3.8-27B. A small extra head scores all options of all questions in one pass, so it never writes free-form text and there is nothing to parse. Cloudflare reports a median latency of about 209 milliseconds in its own test suite.
Model details
Original model Cloudflare/clef
Input Image or text Output Text
Ways to use it Run locally
One reviewed local configuration. Add My computer details to compare them.
BF16 SGLang on NVIDIA · SGLang · The pinned card documents an SGLang deployment using its `dev-clef` Docker image. The 64 GB threshold is a BuildTube estimate for one NVIDIA configuration at 4K.
How this configuration was checked
64 GB VRAM · reviewed estimate Source lists 55 GB (file-size). Estimate uses 54.97 GB exact BF16 model and joint-head bytes, 0.269 GB 4K KV cache (16 full-attention layers of 64 hybrid layers × 4 KV heads × 256 head width × K/V × 2-byte values), 4 GB runtime buffers, and 4.761 GB remaining in the 64 GB VRAM tier.
BF16 · 4,096 token context · source context 262,144 tokens · batch size 1 · concurrency 1 · Single-user 4K comparison using the SGLang Docker image documented by the card.
Source checked 2026-10-21 · revision a30162b99d68a481d8bd75d38df069a0cf83df40. pinned card contains the SGLang `dev-clef` deployment; exact BF16 shard and joint-head bytes plus hybrid attention config supply estimate inputs
Role in this family: Built from another model.
Qwen/Qwen3.8-27B (opens in a new tab)What it does
The model produces a score for each allowed option. A softmax calculation converts those scores into probabilities for each question.
How to use it
Install torch, transformers and the huggingface_hub package, then download the model and load it with the load_release_model helper that ships in the repository. Describe the situation as a state and list your questions with their allowed answers, then call systemone or run a batch. It returns each answer with its confidence and probabilities. You can also serve it with the SGLang Docker image and send requests to /v1/systemone.
Common use cases
- Send a customer message with questions for department, urgency and whether a service is down; get back billing or technical, an urgency level and a true-or-false probability.
- Attach a photo of a receipt and ask whether the total is legible; get back a probability that the answer is true.
- Pass an invoice as JSON with a question about its status; get back probabilities for paid, overdue or draft.
Can I use this?
Requirements The creator tested it with torch 2.11 and transformers 5.10.2 on a single H200 GPU; image and video inputs also need pillow. SGLang serving instructions cover H200, B200 and B300 GPUs. The card does not state minimum hardware. Loading runs the joint_schema_model.py file shipped with the weights.
Caveat All benchmark and latency figures come from Cloudflare’s own internal run, and on several tasks other models in its table score higher. Inputs are capped at 16,384 tokens by default.
Worth knowing Clef scores allowed answers rather than generating free-form text. The default input limit is 16,384 tokens. A smaller, faster variant, Clef-Flash, is published separately. Licensed under Apache License 2.0.
Why it's here
Not trending this week
Last seen on the Hugging Face trending list on 16 October 2026. It stays on BuildTube so you can still find it.
- 2,345 people have liked it on Hugging Face.
- It was downloaded 60,375 times in the last 30 days.
Numbers checked on 8 November 2026; not refreshed since.
Behind it
View model pageBuildTube has not run or verified this project. Everything above is written from what the creator published.
Made this?
Quick questions
- What is Clef?
- The model produces a score for each allowed option.
- What can I use Clef for?
- Send a customer message with questions for department, urgency and whether a service is down; get back billing or technical, an urgency level and a true-or-false probability; Attach a photo of a receipt and ask whether the total is legible; get back a probability that the answer is true; Pass an invoice as JSON with a question about its status; get back probabilities for paid, overdue or draft
- Can I run Clef on my computer?
- It runs on your own hardware with a model runtime. The creator tested it with torch 2.11 and Transformers 5.10.2 on a single H200 GPU; image and video inputs also need pillow. SGLang serving instructions cover H200, B200 and B300 GPUs. The card does not state minimum hardware.
- What do I need to use Clef?
- The documented test setup uses torch 2.11, Transformers 5.10.2 and one H200 GPU. Image and video inputs also need pillow.
- What license does Clef use?
- Licensed under Apache License 2.0.
More from this week
- Idol Turn one picture of an object into a 3D Gaussian splat
Idol turns one photo, render or generated image of an object into a 3D Gaussian splat file.
- P-H3-Image-Comfy Make and edit 8-megapixel images in ComfyUI
Turn short text requests into 8-megapixel images and edit pictures on your own graphics card.
- agent-desktop Let AI agents read and click desktop apps
For developers building computer-use agents: agent-desktop lets an AI agent read and operate desktop apps on Mac, Windows or Linux from a command line.
Explore by topic
Post this
Download share imageFor makers
Made this? Show it’s on BuildTube — add the badge to your README or embed the card on your site.
README badge (Markdown)
[](https://buildtube.app/p/hf-cloudflare-clef/)Embed card (HTML)
<iframe src="https://buildtube.app/embed/hf-cloudflare-clef/" width="560" height="240" style="border:0;max-width:100%" loading="lazy" title="Clef on BuildTube"></iframe>