Clef

Cloudflare's Clef uses a pooled span covering the full option plus a linear head, avoiding the duplicate-first-token problem at the root.

When packing multiple questions, the causal mask needs to be modified: each question can see the full context, but the questions must be isolated from one another; if question two can see the content of question one, the order of asking affects the answers and reproducibility is broken. The author points out that Cloudflare's Clef does not do this isolation, which in his view would cause output drift after reordering.

On the 112 short cases, hosted Jev and PPLX each scored 112, NeoHorse and Clef each scored 104, Flash and Omni each scored 102, Jeff 2B scored 94, Jeff 0.8B scored 77, Laya Typed scored 52, Laya scored 46, and Laya Multilingual scored 25.

The Cloudflare Clef series is entirely open-weight: Clef is based on Qwen3.8-27B, Clef-flash is based on Qwen3.5-9B, and Clef-omni is a 30B mixture-of-experts model (with 3B active parameters).

Clef can be integrated alongside the existing Cloudflare Workers runtime without building a separate observability stack.

Clef collects the behavioral trajectories of Agents at runtime and, through repeatable before-and-after comparisons, makes behavioral drift a visible difference.

In workflow evaluations, clef-omni is lower than Clef on every row: invoice processing exact actions 60.2 versus 64.7, customer service 71.6 versus 76.3; it also trails Jev (61.8 and 76.0).

The Clef API is compatible with Jev, and switching requires only changing the model ID.

On the RAGTruth benchmark, Clef scores 79.4 for detecting RAG output hallucinations, while clef-omni drops to 42.0; therefore clef-omni should not be used as a cheap RAG hallucination checker, and ordinary Clef should be used instead.

The Clef series of decision models supports visual input and can process image data, whereas GPT's corresponding solutions do not always offer this.

Both the Clef model and the Laya model with a Qwen3.5 backbone can be exported to GGUF format and served via llama.cpp, including multimodal decision models capable of understanding images.

In a test on Federal Reserve interest rate questions with market prices hidden, the three models displayed distinct styles: Decisions was the most decisive and willing to contradict the market, Jev was extremely cautious and almost never wavered, and Clef reacted most dramatically and was even misled by SEO spam pages.

Clef is an open-weight model that can be called via Cloudflare Workers AI or run locally with Ollama.

This version supports exporting Clef models based on the Qwen3.5 backbone, as well as Laya models, to GGUF format.

Clef froze training using Coin as the base and simultaneously launched an RL fine-tuning platform.

Clef has a built-in visual encoder, supports a 64K context window, and offers twice the context of comparable models.

Clef is a multimodal model oriented toward audio understanding, capable of processing audio input and generating text descriptions or answering questions related to audio content.

Clef belongs to the Audio-Language Model subcategory.

The noul score type has a value range of 0 to 1 and can be used directly as a decision threshold for business rules.

Both Clef and Clef Flash have multimodal capabilities; requests can contain both images and text, with images shared across all questions and jointly scored with text.