Cloudflare has announced a preview of Clairic, a model it calls a System One extraction model. It runs on Workers AI as @cf/cloudflare/clairic. Cloudflare says it intends to release the weights openly soon, but that has not happened yet.
What it does
Clairic takes a text input, plus up to four optional images, and a map of named fields. Each field gets a type and optional plain-language instructions. The model answers every field in one parallel pass, with no reasoning step and no generated tokens. Cloudflare says that means you pay only for input and typically get a response in under a second.
The field types are text spans, numbers, dates, times, datetimes, durations, lists, estimates and forecasts. Every scalar answer carries a confidence score and p_null, the probability that the value is not in the input. Text answers come back with character offsets, so you can highlight the source passage. If the value is not there, the answer is null rather than a plausible guess, although the response still shows what the model would have picked.
Computing and estimating, not just copying
In Cloudflare's example, a customer email gives a price and a tax amount separately. Clairic returned their sum as the total paid. It also worked out the number of days between two dates in the text. The missing tracking number came back null with low confidence.
Estimate fields return a median and an 80% interval inside bounds you declare. A narrow interval signals high certainty and a wide one signals low certainty, which you can use to trigger another check. Forecast fields return one estimate per future step, each with its own 80% interval.
The evidence Cloudflare offers
For forecasting, Cloudflare ran Clairic daily for 21 days on public TSA checkpoint counts. It reports an average error of 3.3%, with 17 of the 21 actual counts landing inside the 80% interval. Three of the four misses fell on the Columbus Day weekend, because the model forecasts from patterns in the series and does not know about holidays. Cloudflare also says the model is not a replacement for a dedicated time-series model.
On a photographed receipt with imperfect legibility, the model extracted all the requested values with confidence scores. It correctly returned null for the discount. Confidence on the order number and purchase time was lower than on the amounts.
Where it fits in a pipeline
Cloudflare positions Clairic next to Clef, a decision model that picks from options you list, and an LLM that writes the reply. Its demo chains the three on a damaged-product claim. Clairic extracts the facts and estimates repair cost, Clef picks refund or repair against a stated policy, and an LLM writes the response. It frames the idea as the extraction model turning raw input into values, the decision model turning values into decisions, and the LLM turning decisions into words.
The source says Clef and Jev, from TypeSafe, are decision models. Clairic is the first extraction model in this class. The accuracy, speed and cost claims come from Cloudflare's post and are not independently tested.
Top comments (0)