Cloudflare put full open weights for its Clairic extraction models on Hugging Face and X lit up within hours, with the 27B flagship priced at $0.24 per million input tokens and the 9B flash sibling at $0.10. The verified facts underneath the discourse: two downloadable extraction models, Workers AI hosting, and a latency table Cloudflare itself labels self-reported.
Key facts
- Cloudflare's announcement published full open weights for Clairic (27B) and Clairic-flash (9B) on Hugging Face, with both hosted on Workers AI as
@cf/cloudflare/clairicand@cf/cloudflare/clairic-flash. - Published pricing: $0.24 per million input tokens for the 27B Clairic and $0.10 per million input tokens for Clairic-flash, billed on input tokens only, with a promotional discount holding both at $0.10 until December 1, 2026.
- Cloudflare's self-reported latency table shows 358 ms p50 for the 27B model and 155 ms p50 for the 9B model on fresh text fields, measured as complete HTTP responses from a self-hosted server.
- The company's own eval table claims 1,263/1,263 on tracked customer assertions for both new models and 576/576 on synthetic receipt fields across text and images, compared against Liquid AI's D1.2 and Scaledown's Extract.
- The announcement post carries its own caveat: "take these evals as directional guidance since they are self-reported."
What's driving the conversation
The release wave started with Cloudflare's own announcement post and spread fast through the extraction-and-agents corner of X. TeksEdge's thread recapped the open-weights angle, b1rkh0ff pulled out the latency numbers, and replies from agonzalp and musthaveai kept the receipt-extraction use case front and center. The engagement pattern is builder-practical rather than benchmark-war: most replies ask what hardware serves it and how it prices against a general-purpose LLM doing the same job.
The second thread of argument is brand-level. Cloudflare's post opens by calling the company "a neolab" and re-introducing the Workers AI superintelligence team, and part of the discourse is about what that means: a networking company shipping frontier-adjacent models, open weights attached, as a product strategy rather than a research sideline.
The substance
Verified: the weights are downloadable, the Workers AI endpoints are live, and the pricing page numbers are published. The blog post also documents a real architecture change since the preview: parallel slot autoregressive generation, where each extraction field advances with its own end marker instead of a single forward pass, plus new trained heads for estimates, forecasts, durations, and explicit arithmetic. Image reading accuracy moved, by Cloudflare's count, from 207 errors to 0 on its internal benchmark.
Claims, not facts: every number in the eval and latency tables. Cloudflare benchmarked its own models against D1.2 and Scaledown's Extract using internal eval sets, and says so plainly. There is no public extraction benchmark yet, so nobody outside the company can check the 1,263/1,263 line or the 358 ms figure against independent hardware. The model cards confirm the parameter counts and license situation, but performance claims stay vendor-attributed until someone runs them.
Why builders are watching
Three hooks. First, price: $0.10 per million input tokens for extraction-grade output, billed input-only, is a concrete benchmark for what task-specific models cost next to general-purpose API calls. Second, control: invoices, receipts, and contracts are exactly the documents teams hesitate to send to third-party APIs, and open weights make a local extraction path possible. Third, the December 1 pricing deadline: the promotional rate expires then, the 27B endpoint returns to $0.24, and anyone who built on the preview endpoint inherits the new model automatically, so the swap lands whether or not you asked for it. This is the second task-specific open-weight model family DeAI has covered after Fastino's GLiNER2.5, and it follows Cloudflare's own SGLng serving-stack incident, covered here previously, which is a reminder that self-hosting the weights is also the option that takes the serving stack out of someone else's hands. The open-weight versus open-source distinction applies cleanly here: weights are open, the serving product is Cloudflare's.
Questions
- What did Cloudflare release as open weights?
- Full open weights for Clairic, a 27B-parameter document-extraction model, and Clairic-flash, a 9B variant, published on Hugging Face and hosted on Workers AI. Both pull dates, numbers, tables, and confidence spans out of text, images, and PDFs.
- What does Clairic cost on Workers AI?
- Cloudflare's published rates are $0.24 per million input tokens for the 27B Clairic and $0.10 per million input tokens for 9B Clairic-flash. A promotional discount holds both at $0.10 until December 1, 2026, and billing covers input tokens only.
- What's disputed about the Clairic release?
- The accuracy and latency tables, including 358 ms p50 for Clairic and 155 ms p50 for Clairic-flash, are Cloudflare's internal evaluations. Cloudflare itself says to take them as directional since they are self-reported, and no independent extraction benchmark exists yet.
- Can Clairic run locally?
- The weights are downloadable from Hugging Face, and the model card paths cover local runs with Transformers and vLLM. Cloudflare's published latency figures describe its self-hosted test server, not arbitrary local hardware.
Sources
About DeAI
DeAI is an independent publication covering open-weight AI models, private inference, and decentralized infrastructure — the tools for running AI you actually control. We test providers on price, privacy, and refusal behavior and publish the numbers, not the vibes. DeAI is powered by Morpheus (mor.org), a decentralized inference marketplace, and covers it on the same terms as every other provider.
Powered by Morpheus and StrandCMS
Morpheus is a decentralized inference marketplace, covered on the same terms as every other provider — we rank it wherever the data lands. StrandCMS is the open-source, agent-first framework this site is built on.
