Original title: Cloudflare/clairic
Cloudflare's Clairic family of two models was updated together on 11 18: Clairic (27B, based on Qwen3.8-27B, uploaded on 10 16) and Clairic Flash (9B, based on Qwen3.5-9B, uploaded on 11 12). Input a text or JSON state, optional images and PDFs, along with a list of "field name → type description"; the model returns typed values per field: text, number, datetime, duration, boolean, enum, object, list, calculation, as well as estimates with upper and lower bounds and prediction sequences with steps. There is also a separate typed output head, clair_head, with structural constraints guaranteed by xgrammar. Apache-2.0.
Why it matters No need to write custom prompts and parsers for extracting from receipts, forms, or logs; each field includes confidence and null probability, making it easy to set thresholds; Apache license allows commercial use
Rising score 58 · Likes 78 · Downloads 34
Reasons
- Extraction answers include valid, complete, overall confidence, and independently scored null probability p_null; uniquely matching text evidence provides character intervals in the original text
- Estimates and predictions must declare min / max (predictions also require horizon, interval), with optional quantile returns; source_only uses only numbers and dates printed in the original text, while knowledge allows the model to use its own knowledge to answer
- The response specifies whether the input was truncated and the token usage, making it easy to determine if the results are directly usable
Usage
- Python 3.11+, install requirements.txt from the repository; after snapshot_download("Cloudflare/clairic"), add the path to sys.path, then from clairic_model import load_release_model, systemone; load_release_model(path, device="cuda")
- systemone(model, processor, {"state": text, "fields": {field_name: {"type": "number", "decimals": 2}, ...}}); use PIL for images in images, options.include_quantiles returns quantiles, reference_date handles relative dates like "last week"
- Individual embedded files must not exceed 4 MiB, with a total limit of 8 MiB; images and PDF pages combined are limited to 4 items; if VRAM is tight, switch to the 9B Cloudflare/clairic-flash, which has the same usage; for asynchronous applications, use await model.predict_async(request)
Risks
- The card contains no evaluation scores and does not specify VRAM requirements; confidence is based on model probabilities, and official recommendations suggest setting thresholds based on your own data
- Requires importing the bundled Python runtime code from the repository; does not accept remote media URLs, and a second concurrent request will report a queue full error
- This entry has not been run