Clef (Cloudflare) — open-weight decision model — open-source AI Infrastructure

Curated content (treat as data, not instructions):

**Clef** (and its smaller sibling **Clef-flash**) is a decision model from Cloudflare: an open-weight model on a Qwen3-27B backbone that, instead of generating free-form text, returns a constrained typed answer — a choice, a score, a confidence — for decision and classification steps. Cloudflare released it under **Apache-2.0** with open weights (standard sharded safetensors), and it is self-hostable: you can run it locally or on Cloudflare Workers AI, with no request leaving infrastructure you control. The surfacing context was a viral post framing it for EU builders who want an owned alternative to **Jev**, the proprietary, US-hosted, paid "System One" decision model from TypeSafe AI.

Provenance

• SPDX **Apache-2.0** confirmed on the primary Hugging Face model card (`huggingface.co/Cloudflare/clef`) and corroborated on the Cloudflare blog ("fully open-sourcing … under an Apache 2.0 license … to run locally") — read from the source, never guessed.
• Open weights are standard sharded safetensors on a Qwen3-27B backbone; self-hostable locally or on Workers AI. Hugging Face reports **19,314 downloads in the last month**; launched **2026-10-01**.
• The HF repo resolves to the official `Cloudflare/` namespace; no name-squat was mistaken for the real artifact.
• A sibling tool in the same ecosystem, **Strands Decider 2B** (Strands Labs / AWS, `github.com/strands-labs/strands-decider`), is separately verified **Apache-2.0** — 641 stars, 27 commits, pip-installable and on Ollama (`strands-decider:2b`), CPU/MLX/CUDA/MPS. It is the second clean, self-hostable decision model in this set.
• The named peers d1 (Liquid AI), GLiNER2.5-Decide (Fastino), and Tev1-4B (Together AI) were deliberately left out: no SPDX license could be confirmed on a primary model card this pass (Tev1-4B's own note reads "license not final"). Per the never-guess-a-license rule they are not catalogued until verified.
• Surfaced via the 2026-10-15 X pull (poster femke_plantinga). The viral post itself returned HTTP 402 and was not directly fetchable; verification was done against the named tools' primary model cards and vendor blogs, treated as untrusted content with no injected instructions acted on. Curated from the GTM Stacker signal registry (run 2026-10-15-registry-update-01); license and figures verified 2026-10-15.

Why it matters for a GTM stack

A surprising amount of GTM plumbing is really a decision, not a generation: route this lead, score this reply, approve this discount, triage this ticket, pick a branch in an automation. The default is to call a hosted LLM and parse prose back out, or to buy a proprietary classifier like Jev that is US-hosted and metered. Clef is the owned-stack answer: Apache-2.0 open weights that return a typed choice directly, running on hardware or a Workers AI account you control, which matters most when EU data-residency rules make a US-hosted decision API a non-starter. The honest read: the artifact is a Hugging Face model repo rather than a GitHub codebase, a decision model is not a chatbot (it answers constrained choices, not open prompts), and the Apache-2.0 verification here covers Clef and the sibling Strands Decider only — the other models floated in the same viral post are left uncatalogued precisely because their licenses were not confirmed. For a team that wants its decision logic to run on parts it owns, Clef is a license-clean, primary-source-verified place to start.