Cloudflare/Clef Flash 9B

• License: Apache-2.0
• Release date: 2026-10-01 (public announcement)
• Modalities: Input: text, images; Output: typed decisions (choice, score, noul)
• Original model: Cloudflare/clef-flash

Description

Clef Flash combines a Qwen3.5-9B multimodal backbone with a trained joint schema head. It reads a state and typed questions together, scoring every allowed option in a single forward pass instead of generating free-form text.

Run

self serve clef-flash:9b

Here is a link to the API docs: Decision API.

Example Request

With the server running, open another terminal and send a request:

curl http://localhost:8080/v1/systemone \
  -H 'Content-Type: application/json' \
  -d '{
    "state": "Checkout is failing and customers cannot place orders.",
    "questions": {
      "department": {
        "type": "choice",
        "instructions": "Which team should handle this ticket?",
        "criteria": {
          "billing": "Payments and refunds",
          "technical": "Bugs and outages"
        }
      }
    }
  }'

The server returns a JSON decision for the department question.

Engine Parameters

Served by the native clef engine through the clef adapter.

Parameter	Default	What it controls
context_size	16384	Token budget shared by state, schema and images; accepts 512 to 262144.
threads	Hardware concurrency (at least 1)	CPU workers; accepts 1 to 1024.
gpu_layers	-1	GPU-offloaded backbone layers; -1 means all, 0 means CPU; accepts -1 to 10000. CPU device selection forces zero offload.

Override startup settings with --set <parameter>=<value>.

Quantizations and Files

Default variant: q4. Sizes and SHA-256 hashes are pinned in the model registry.

q4 (default)

File	Role	Size	SHA-256	Origin
Cloudflare_clef-flash-Q4_K_M.gguf	model	5.63 GiB (6040541344 bytes)	cf747351878a0e569fd9d6681395bd723e111f63a4f6fe6e26e7563dfdfa9b0a	bartowski/Cloudflare_clef-flash-GGUF
mmproj-Cloudflare_clef-flash-f16.gguf	mmproj	875.6 MiB (918165984 bytes)	c8a6e9e00e1baf6433c4874dbc73bed2a56e887bcbb582fffcb00b533bc638dd	bartowski/Cloudflare_clef-flash-GGUF
joint_head.safetensors	head	232.3 MiB (243538016 bytes)	19cdcec8c81dc9212be320fff47462ab342fbc1278be4368fb3da71241cf5ba0	Cloudflare/clef-flash

Total download: 6.52 GiB (7002756992 bytes).

Notes

• On 2026-10-13, the Q4 backbone, projector and joint head were pulled and verified locally. CPU text decision probes passed routing, option-order swaps, true/false probabilities and score ordering after fixing decoder-state extraction before the GGUF classifier projection.
• Vision accuracy is currently not validated: the CPU solid-color acceptance test classified red and green images as blue. Image decoding and malformed/excess-image rejection work, but those checks do not establish correct image decisions. Do not rely on this entry for visual decisions until this failure is resolved.
• Up to 32 options per question and 4 images per request. Images share the default 16384-token context with the state and schema; an oversized state is truncated.
• Send images in the images array as URLs or JPEG, PNG or WebP data URIs. Refer to them by position ("the first image"); all questions are answered together.
• The confidence of a choice is the probability of the chosen option. No free-form text generation or video-input API is exposed here, even though upstream supports video.
• The backbone, vision projector and joint head are all required. Download size is not the total runtime memory requirement.
• Origin links use main because these registry entries pin hashes, not repository revisions.

Quality benchmarks

Decision Index: #54 (47.61)

Cloudflare/clef-flash

Published result for the named upstream model and evaluated configuration; checkpoint revision and local quantization equivalence are not guaranteed.

Evaluated: clef-flash

Edition: Decision Index 0.3.2

Checked 2026-10-13

JevBench: #44 (36.64)

Cloudflare/clef-flash

Published result for the named upstream model and evaluated configuration; checkpoint revision and local quantization equivalence are not guaranteed.

Evaluated: Clef-Flash (Cloudflare, Qwen3.5-9B post-train with a joint schema head, multimodal, measured on text)

Edition: v1.6.1

Checkpoint: Cloudflare/clef-flash@17f0b0ad64efb65d273590632833508766b2aae6

Checked 2026-10-13

Benchmarks

No submitted benchmark reports yet.