What Cloudflare released

On 18 November 2026 Cloudflare announced a redesigned Clairic and a new, smaller Clairic-flash, and put both on Hugging Face as open weights. Both models do structured extraction. You hand over some text, a PDF or a photo along with a list of named fields and their types, and each model returns one typed answer per field. The Clairic model card lists a licence of Apache-2.0 and a base model of Qwen3.8-27B. The Clairic-flash card lists Apache-2.0 and Qwen3.5-9B. Apache 2.0 allows commercial use, modification and redistribution, which matters for anyone who wants to fine-tune on their own documents or ship the model inside a product.

Clairic first appeared on 20 October as a hosted preview. Help Net Security reported then that Cloudflare intended to publish the weights later. The November release delivers those weights, adds PDF input and moves the 9B model to a new endpoint name. RuntimeWire covered the release on 18 November and confirmed the pricing below.

How a typed extraction model works

Most teams that pull data out of invoices or claim forms prompt a general chat model to "return JSON" and then validate what comes back. Clairic moves the schema into the model interface. According to the model card, a request carries a state (the source text or JSON), up to four images or rendered PDF pages, and a fields map. Each field has a type such as text, number, date, duration, boolean, enum, list, compute, estimate or forecast. Options include nullable, which defaults to true, and source_only, which restricts answers to evidence found in the input.

The response is where Clairic differs most from a prompted chat model. Every scalar answer carries a confidence score and a separate p_null, which is the model's estimated probability that the value is absent from the input. Text answers found uniquely in the source come with character spans, so a reviewer can highlight exactly where an invoice number came from. Cloudflare's blog says ordinary text and number fields now use "parallel slot autoregressive generation". Each field generates its own tokens and stops at its own end marker, and all fields advance together in one serving batch. Cloudflare trained separate heads for estimates, forecasts, durations, calculated dates and explicit arithmetic. The repository ships those heads as clair_head.safetensors alongside the main weights.

Cloudflare made a useful admission during the preview: "The architecture makes every answer well formed, not necessarily correct," as quoted by Help Net Security. The schema guarantees a date field will hold a date. Nothing guarantees the date is the right one, so the confidence and p_null scores are what you build review thresholds on.

The numbers, and how far to trust them

The Cloudflare blog publishes internal results. On 336 fresh text fields, Clairic got 335 right and Clairic-flash got 328, against 295 for the October preview. On 226 mixed-input fields the scores were 221 for Clairic and 213 for Clairic-flash, while a comparison pipeline Cloudflare labels "D1.2 + OCR" managed 116. Median latency on a self-hosted server ranged from 358 to 691 ms for Clairic and from 155 to 393 ms for Clairic-flash, depending on the test set.

Cloudflare wrote every one of those evals, and the blog says no public extraction benchmark exists yet and calls its comparisons directional. Neither model card reports any benchmark. Treat the figures as a reason to test the models yourself. They don't show how the models will handle your documents.

The preview write-up also listed limits that Cloudflare hasn't said are fixed. Help Net Security reported that one- and two-step calculations looked reliable. Longer chains of dependent arithmetic and sums across many table rows didn't, and Cloudflare advised doing that maths in code. Prediction intervals could also come out too narrow on unfamiliar data.

Running the models yourself

The model card's supported path is Python 3.11 or later with PyTorch and Transformers, using the helper code that ships in the repository:

from huggingface_hub import snapshot_download
path = snapshot_download("Cloudflare/clairic")
sys.path.insert(0, path)
from clairic_model import load_release_model, systemone
model, processor = load_release_model(path, device="cuda")

You then call systemone(model, processor, request) and read response["answers"]. The Clairic-flash card adds xgrammar for schema constraints. Hugging Face also shows auto-generated vllm serve and SGLang commands on the Clairic card, but the card itself only documents the PyTorch path. Expect the typed heads to need the bundled helper code.

Neither card gives VRAM figures. Our own rough arithmetic, at two bytes per parameter in BF16, puts the weights alone near 56 GB for Clairic (Hugging Face metadata says 28B parameters) and near 20 GB for Clairic-flash (metadata says 10B). Activations and image tokens come on top of that. Clairic-flash at full precision should fit on a single 24 GB card with little headroom. Clairic at full precision needs a large workstation GPU, two cards or a quantised build. Nobody has published a quantised build with the typed heads intact yet, as far as we can find.

Both cards list the same input limits: 4 MiB per embedded file, 8 MiB combined, at most four images or PDF pages per request, and no remote URLs. Multi-page documents therefore need splitting before they go in.

Hosted costs and a billing catch

The Workers AI docs for Clairic-flash list US$0.10 per million input tokens, with no charge for output tokens. Clairic costs US$0.24 per million input tokens, and both models cost US$0.10 until 1 December 2026, according to Cloudflare. The context window is 65,536 tokens.

The docs also explain how passes work. Each text field gets its own pass, and other field types share passes where they fit. The source text and every image are billed again on each pass. A request for ten free-text fields over a long contract can therefore cost close to ten times the input tokens of a single read. The usage.passes value in each response tells you how many passes ran, and it's worth logging from day one.

What NZ builders could do with it

Accounts payable is the obvious use: supplier invoices, receipts and expense claims, where a reviewer only sees fields with low confidence or a high p_null. The character spans make that review screen simple to build. A self-hosted Clairic-flash keeps documents on hardware you control. That helps when clients, such as health providers, councils or law firms, won't accept documents leaving the country, since Workers AI doesn't let you choose a New Zealand inference region for these models. The unit and reference_date options in the docs are worth testing with NZD amounts and day-first dates before you trust the output on local paperwork.

A sensible trial is a few hundred of your own historical documents with known answers. Score Clairic-flash against whatever pipeline you run now, then pick confidence thresholds from the results. If you build something with it, we'd like to see it on the ASAI projects page.

Sources

Build with people working on the same things

Join the Auckland Society for AI: NZ$599 a year with a ForgeBox and the members' projects area. Our Auckland meetups are open to everyone.

Become a member