Cloudflare Releases Clef-omni: An Apache 2.0 Multimodal Model That Returns Decisions, Not Text

The Three-Question Framework

What Happened?

Cloudflare published the weights of Clef-omni on Hugging Face under Apache 2.0. The model card says it is post-trained from Qwen/Qwen3-Omni-30B-A3B-Instruct, a 30B-total, 3B-active mixture-of-experts model. It accepts text, JSON, images, audio and video, and returns a probability for each allowed option of each question in a single forward pass, with no free-form text generation.

Why It Matters

The design targets classification and routing jobs such as content checks, where a yes/no or multiple-choice answer is needed quickly and without parsing model prose. The card says the backbone needs about 64 GB of GPU memory in bfloat16. Its own table shows the trade-off: strong on MMLU and BANKING77, weak on GPQA Diamond and RAGTruth.

What Could Happen Next

Developers get an open model for structured decisions that they can self-host. Scores come from Cloudflare's internal "Decision Index 0.2.1" run, so independent tests are the next thing to watch. It sits beside Cloudflare's dense Clef and Clef-Flash image and video variants.

Executive Key Takeaways

Base model: Qwen3-Omni-30B-A3B-Instruct (30B total, 3B active); Hugging Face metadata lists 35B parameters in BF16.

Output is a probability per allowed option, with no text generation and no output parsing.

Licence Apache 2.0, following the base model; default maximum input length 64,000 tokens.

Self-reported scores: MMLU 92.7, ARC-Easy 99.8, BANKING77 94.8, GPQA Diamond 47.4.

Tested on one H200; about 64 GB of GPU memory needed in bfloat16.

Everything here comes from the Hugging Face model card for Cloudflare/clef-omni, which we read directly. The scores are Cloudflare's own, from an internal Decision Index 0.2.1 run, and have not been independently reproduced.

The model is a decision model. Instead of writing an answer, it scores each allowed option for each question in a single forward pass. The card says there is no free-form generation and no output parsing, which removes a common failure point in pipelines that ask a chat model to return JSON.

Inputs cover text, JSON, images, audio and video. The card lists a default maximum length of 64,000 tokens in its encoding helper. It was tested with torch 2.11 and transformers 5.10.2 on a single H200.

Reported results: MMLU 92.7, ARC-Easy 99.8, HellaSwag 99.1, CRUXEval 88.8 and BANKING77 94.8, which the card marks as leading. On harder reasoning it trails: GPQA Diamond 47.4 against 78.3 for a model the card calls Jev, MMLU-Pro 70.6 against 82.7, and RAGTruth 42.0 against 79.4 for Clef. On an invoice-processing workflow test its exact-action accuracy was 60.2 against 64.7 for Clef.

Clef and Clef-Flash are dense image and video variants in the same family. Clef-omni is the one that adds audio and uses the mixture-of-experts backbone.

The card shows 297 downloads in the last month, so adoption is small so far.

The Three-Question Framework

What Happened?

Why It Matters

What Could Happen Next

Executive Key Takeaways

Primary Sources & Citations