· Nov 18, 2026
Cloudflare opens weights for updated Clairic and Clairic-flash
Markdown · JSON · Original source
says it redesigned Clairic around extraction from unstructured data, including images and PDFs, and released an 9-billion-parameter version, Clairic-flash. The company says the redesign uses parallel slot autoregressive generation for ordinary text and number fields, while special heads handle estimates, forecasts, durations, calculated dates, and explicit arithmetic. Cloudflare reports improved results on its internal extraction evaluations and says image-reading errors fell from 207 to 0 on internal . It cautions that its comparisons are directional because they are self-reported, and notes that public extraction benchmarks do not yet exist. The company also reports sub-second HTTP response times in tests on a self-hosted server; those measurements do not establish performance for other deployments. Both models are available on Workers AI and as open weights on . Cloudflare frames them as smaller, faster, cheaper models for specific tasks that could supplement expensive use, while inviting others to experiment with the releases.
Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.
Original source
Blog · Article
Source details
Original source date: Nov 18, 2026 (2026-11-18T22:00:00.000Z).
Date basis: Date stated by the publisher.
Added to Escape Meridian: 2026-11-18T22:17:24.620Z. First observed: 2026-11-18T22:10:12.711Z. Retrieved: 2026-11-18T22:16:22.829Z.
Referenced by
Recorded citations in public readings. A reference alone does not establish agreement or influence.
No public references recorded here yet.
Limits
says there are no public extraction yet and describes its comparative evaluations as directional and self-reported. The reported tests were measured from a self-hosted server; the source does not establish that the results apply to other deployment conditions. The reported evaluation results do not establish performance on extraction workloads beyond the tested cases.
Open question
How would Clairic and Clairic-flash perform on public extraction beyond the internal evaluations reports?
Source-stated
Cloudflare presents the releases as an updated Clairic focused on extraction and a lighter, faster Clairic-flash model; both are available as open weights.
We’re introducing an updated version of Clairic. Since the preview, we’ve improved extraction from complex documents including images and PDFs (new!)
Source-stated
Cloudflare says it redesigned Clairic’s architecture for extraction, using parallel slot autoregressive generation for ordinary text and number fields and special heads for estimates, forecasts, durations, calculated dates, and arithmetic.
For ordinary text and number extraction fields, we now use parallel slot autoregressive generation instead of a single forward pass.
Source-stated
Cloudflare reports improved internal evaluation results and says image-reading errors declined from 207 to 0 on its internal benchmarks.
Due to Clairic’s new architecture, image reading accuracy has improved from 207 errors to 0 on our internal benchmarks, explicitly covering text and images.
Source-stated
Cloudflare says its evaluation comparisons are directional and self-reported, and that public extraction benchmarks are not available yet.
There are no public extraction benchmarks yet, but we ran our internal evals for other emerging extraction models on the market
Source-stated
Cloudflare reports measuring full HTTP response latency on a self-hosted server, with results under one second in its tests.
We ran latency tests for Clairic and Clairic-flash, measuring complete HTTP response from a self-hosted server, including all fields in the request. Even with the redesigned architecture, you can expect responses in under one second
Source-stated
Cloudflare says the models are intended to supplement expensive LLM usage with models built for specific tasks.
Our mission with these System One models is to help people augment their expensive LLM usage with smaller, faster, cheaper models that are purpose-built for a specific task.
Source and contributor context
Cloudflare · Publisher
The publisher is identified here. No individual author has been inferred.
Source media
Media declared by the original publisher; availability checked separately from the editorial reading.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Open weights
A model whose trained numbers are published so others can download, run and inspect it; the license and training data may still be restricted.
Benchmark
A defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Large language model
A language model with many learned numerical settings, used to process and generate language.
Latency
The delay between sending a request and receiving the result.
Cloudflare
Cloudflare provides networking, security and computing services that help run websites and applications.
Hugging Face
Hugging Face hosts models, datasets and software that let people share, adapt and inspect AI systems.
Editorial provenance
Version 1. Publication ID: dc725439-ffc1-4f17-823b-0a1a66340fa0.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-18T22:17:24.620Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: e8d390fc0f0f062ebac66bcf29a90755c984f5fbe84a698288b66e13afce2210. Source text hash: cf031136b956fa415145926cee271cfb1a4e630959b9a949691f7a10c6b7a4c5.
Revision and source-check history
Version 1 · 2026-11-18T22:17:24.620Z
Cloudflare opens weights for updated Clairic and Clairic-flash
says it redesigned Clairic around extraction from unstructured data, including images and PDFs, and released an 9-billion-parameter version, Clairic-flash. The company says the redesign uses parallel slot autoregressive generation for ordinary text and number fields, while special heads handle estimates, forecasts, durations, calculated dates, and explicit arithmetic. Cloudflare reports improved results on its internal extraction evaluations and says image-reading errors fell from 207 to 0 on internal . It cautions that its comparisons are directional because they are self-reported, and notes that public extraction benchmarks do not yet exist. The company also reports sub-second HTTP response times in tests on a self-hosted server; those measurements do not establish performance for other deployments. Both models are available on Workers AI and as open weights on . Cloudflare frames them as smaller, faster, cheaper models for specific tasks that could supplement expensive use, while inviting others to experiment with the releases.
Original source
Blog · Article
Source details
Original source date: Nov 18, 2026 (2026-11-18T22:00:00.000Z).
Date basis: Date stated by the publisher.
Added to Escape Meridian: 2026-11-18T22:17:24.620Z. First observed: 2026-11-18T22:10:12.711Z. Retrieved: 2026-11-18T22:16:22.829Z.
Limits
says there are no public extraction yet and describes its comparative evaluations as directional and self-reported. The reported tests were measured from a self-hosted server; the source does not establish that the results apply to other deployment conditions. The reported evaluation results do not establish performance on extraction workloads beyond the tested cases.
Open question
How would Clairic and Clairic-flash perform on public extraction beyond the internal evaluations reports?
Source-stated
Cloudflare presents the releases as an updated Clairic focused on extraction and a lighter, faster Clairic-flash model; both are available as open weights.
We’re introducing an updated version of Clairic. Since the preview, we’ve improved extraction from complex documents including images and PDFs (new!)
Source-stated
Cloudflare says it redesigned Clairic’s architecture for extraction, using parallel slot autoregressive generation for ordinary text and number fields and special heads for estimates, forecasts, durations, calculated dates, and arithmetic.
For ordinary text and number extraction fields, we now use parallel slot autoregressive generation instead of a single forward pass.
Source-stated
Cloudflare reports improved internal evaluation results and says image-reading errors declined from 207 to 0 on its internal benchmarks.
Due to Clairic’s new architecture, image reading accuracy has improved from 207 errors to 0 on our internal benchmarks, explicitly covering text and images.
Source-stated
Cloudflare says its evaluation comparisons are directional and self-reported, and that public extraction benchmarks are not available yet.
There are no public extraction benchmarks yet, but we ran our internal evals for other emerging extraction models on the market
Source-stated
Cloudflare reports measuring full HTTP response latency on a self-hosted server, with results under one second in its tests.
We ran latency tests for Clairic and Clairic-flash, measuring complete HTTP response from a self-hosted server, including all fields in the request. Even with the redesigned architecture, you can expect responses in under one second
Source-stated
Cloudflare says the models are intended to supplement expensive LLM usage with models built for specific tasks.
Our mission with these System One models is to help people augment their expensive LLM usage with smaller, faster, cheaper models that are purpose-built for a specific task.
Source and contributor context
Cloudflare · Publisher
The publisher is identified here. No individual author has been inferred.
Source media
Media declared by the original publisher; availability checked separately from the editorial reading.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Open weights
A model whose trained numbers are published so others can download, run and inspect it; the license and training data may still be restricted.
Benchmark
A defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Large language model
A language model with many learned numerical settings, used to process and generate language.
Latency
The delay between sending a request and receiving the result.
Cloudflare
Cloudflare provides networking, security and computing services that help run websites and applications.
Hugging Face
Hugging Face hosts models, datasets and software that let people share, adapt and inspect AI systems.
Editorial provenance
Version 1. Publication ID: dc725439-ffc1-4f17-823b-0a1a66340fa0.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-18T22:17:24.620Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: e8d390fc0f0f062ebac66bcf29a90755c984f5fbe84a698288b66e13afce2210. Source text hash: cf031136b956fa415145926cee271cfb1a4e630959b9a949691f7a10c6b7a4c5.
uncertain · 2026-11-18T22:17:24.620Z · anthropic/claude-opus-5.5
Publication: dc725439-ffc1-4f17-823b-0a1a66340fa0
attribution — pass: The reading attributes the redesign, evaluation results, latency figures and mission framing to Cloudflare throughout. The self-reported caveat is presented as Cloudflare's own statement. No source assertion is presented as established fact.
take these evals as directional guidance since they are self-reported
summary — uncertain: The summary is accurate but leaves central terms unexplained for a non-specialist. 'Open-weight', the headline development, is not glossed. 'Parallel slot autoregressive generation' and 'special heads' are also not glossed, and no term-library tooltip is indicated. Smallest repair: briefly say that open weights mean anyone can download and run the model, and add tooltip entries for the other two terms. The reading also drops Cloudflare's framing that it has 'become a neolab' with a 'superintelligence team' building frontier models. That framing is an institutional claim relevant to this stream; consider one attributed sentence.
Cloudflare has become a neolab. We’re building our own frontier AI models for our customers and the open-source community.
date — pass: originalPublished (2026-11-18T22:00Z) comes from the publisher page and matches the feed entry. Nothing indicates an earlier first publication of this post. The draft makes no conflicting date claim.
support — pass: Each claim matches the source. The figure of 207 errors falling to 0 is consistent with the receipt eval rising from 369/576 to 576/576. The sub-second latency claim is consistent with all reported p95 values being under 1000 ms on a self-hosted server. The limitations and the question about public benchmarks follow from the source's statement that no public benchmarks exist. The reading adds no unsupported benefits or costs.
There are no public extraction benchmarks yet, but we ran our internal evals for other emerging extraction models on the market
media — pass: The selected media is the verified direct link to the original Cloudflare blog post, which is the most direct way to access the source.
Challenge or correct this reading
Send a correction or an opposing source with this reading’s URL. Disagreement and uncertainty can remain unresolved.
Cloudflare provides networking, security and computing services that help run websites and applications.
Explore CloudflareCloudflare provides networking, security and computing services that help run websites and applications.
Explore CloudflareA model whose trained numbers are published so others can download, run and inspect it; the license and training data may still be restricted.
Definition referenceA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referenceHugging Face hosts models, datasets and software that let people share, adapt and inspect AI systems.
Explore Hugging FaceA language model with many learned numerical settings, used to process and generate language.
Definition referenceCloudflare provides networking, security and computing services that help run websites and applications.
Explore CloudflareCloudflare provides networking, security and computing services that help run websites and applications.
Explore CloudflareA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referenceThe delay between sending a request and receiving the result.
Definition referenceA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referenceCloudflare provides networking, security and computing services that help run websites and applications.
Explore CloudflareCloudflare provides networking, security and computing services that help run websites and applications.
Explore CloudflareA model whose trained numbers are published so others can download, run and inspect it; the license and training data may still be restricted.
Definition referenceA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referenceHugging Face hosts models, datasets and software that let people share, adapt and inspect AI systems.
Explore Hugging FaceA language model with many learned numerical settings, used to process and generate language.
Definition referenceCloudflare provides networking, security and computing services that help run websites and applications.
Explore CloudflareCloudflare provides networking, security and computing services that help run websites and applications.
Explore CloudflareA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referenceThe delay between sending a request and receiving the result.
Definition referenceA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referenceCloudflare provides networking, security and computing services that help run websites and applications.
Explore Cloudflare