Skip to content

Cloudflare open-weights Clairic and Clairic-flash

Cloudflare says its Clairic extraction model has left preview and a faster-sounding sibling, Clairic-flash, has joined Workers AI, with open weights for both.

Draftship Editorial3 min read

Cloudflare takes the wraps off two extraction models

On November 18, Cloudflare published a post announcing that an updated Clairic and a new Clairic-flash are now open-weight models. The Clairic open-weight release marks the model's exit from preview, while Clairic-flash arrives on Workers AI for the first time. Cloudflare credits both to its new Workers AI Superintelligence team, according to the short announcement post.

That is roughly the whole of the announcement. The text supplied is mostly a teaser line and site navigation, so the confirmed facts are few and the detail is thin.

What the source confirms

Stripping it back to statements Cloudflare actually makes:

  • Clairic has graduated from preview.
  • Clairic-flash is new and is now part of the line-up on Workers AI.
  • Both are described as extraction models.
  • Both are being released as open-weight models.
  • The work comes from a team Cloudflare calls Workers AI Superintelligence.
  • The post is dated November 18.

Nothing else is stated. There are no parameter counts, no context window, no benchmark results, no licence, no price, and no link to where weights can be downloaded.

What open weights and extraction likely mean for a builder

This part is inference, not something the post says.

Open weights usually means the trained parameters are published so others can download, self-host or fine-tune them, rather than only calling a hosted endpoint. If that holds here, it suggests you would not be tied to one provider's inference API for these models, though the source does not confirm that weights are downloadable, where from, or under what terms.

Extraction models are typically used to pull structured fields out of unstructured text, such as turning a wall of prose into a set of named values. The source does not describe what Clairic extracts or how output is shaped, so treat the general definition as context rather than a specification.

The two-name structure, a standard model plus a flash variant, likely signals a speed and cost trade-off rather than a capability split. That reading is a guess. The post does not compare the two.

The preview exit matters on its own. A model that was previously in preview now sits outside it, which usually means fewer warnings attached to production use, but Cloudflare does not state any service level, deprecation policy or stability guarantee here.

Limits and open questions

There is a lot a builder would need before planning around either model. The announcement does not give:

  • the licence terms attached to the weights
  • model size or architecture
  • context length and input limits
  • throughput, latency or accuracy figures
  • pricing for hosted calls on Workers AI
  • which regions or plans can access Clairic-flash
  • whether the previous Clairic behaves differently from the updated one
  • any migration notes for teams already using the preview version

The post also does not say what changed inside the updated Clairic, so anyone who built on the preview has no stated basis for judging whether their results will shift.

What this means if you build apps with AI

For anyone shipping an app that needs to turn messy text into clean fields, a second extraction model with open weights is worth watching, mainly because it may widen your hosting options later. Until licence and pricing details appear, the practical move is to keep extraction behind an interface you can swap, so a model change is a configuration edit rather than a rewrite.

If you are assembling that kind of app without writing the plumbing yourself, Draftship's documentation on adding AI to your app covers how model calls fit into a project. Draftship is not affiliated with Cloudflare and does not claim support for Clairic or Clairic-flash.

Primary sources

This article was drafted with AI assistance from the primary sources listed below and checked automatically against them. An editor hasn't reviewed it yet.