deniz.in

Markets

Weather

Loading weather

· via Cloudflare blog

Cloudflare open-weights updated Clairic and faster Clairic-flash extraction models

Cloudflare's newly named Workers AI Superintelligence team has released an updated Clairic extraction model and a lighter 9B Clairic-flash, with open weights on Hugging Face.

Cloudflare open-weights updated Clairic and faster Clairic-flash extraction models

Cloudflare has released two open-weight extraction models: an updated Clairic and a new, smaller Clairic-flash. The announcement doubles as an introduction to the team behind them, the Workers AI Superintelligence group, which Cloudflare describes as its frontier AI research arm and the outfit responsible for Clef and Clairic.

According to the Cloudflare blog, both models are served on Workers AI under the endpoints @cf/cloudflare/clairic and @cf/cloudflare/clairic-flash, and the full weights are published on Hugging Face.

A rebuilt Clairic

Feedback from the preview release pushed the team to sharpen Clairic's focus on extraction from unstructured data, with use cases ranging from receipt photos to support tickets. The redesign replaces a single forward pass for ordinary text and number fields with what Cloudflare calls parallel slot autoregressive generation: each field picks its next token based on what it has already produced and stops at its own end marker, while all fields advance together in the serving batch. The company says the accuracy gains outweigh the small latency cost.

For computational work such as estimates, forecasts, durations, calculated dates and explicit arithmetic, the models rely on specially trained heads. The typed API from the preview carries over, so responses still include confidence scores, character spans and probabilities.

The model family now also reads PDFs. A single request can include up to four images or PDF pages in total, and line items can return as rows that keep each description, quantity and price together, even when they sit on different pages. Cloudflare reports that image reading errors on its internal benchmarks fell from 207 to zero with the new architecture.

Self-reported benchmarks

Cloudflare notes that no public extraction benchmarks exist, so it ran internal evaluations against other emerging extraction models on the market: Liquid AI's D1.2, pipelined with Apple Vision for optical character recognition, and ScaleDown's Extract. The company cautions that the results are directional guidance because they are self-reported.

On tracked customer assertions covering 139 request shapes, both new models scored a perfect 1,263 out of 1,263, against 1,033 for Clairic-preview, 973 for D1.2 with OCR and 653 for ScaleDown. On synthetic receipt fields in text and images, Clairic and Clairic-flash both reached 576/576 versus 369/576 for the preview model.

Latency tests, measured as complete HTTP responses from a self-hosted server including all requested fields, put Clairic-flash at a median of 155 ms on fresh text fields, with 328 ms at the 95th percentile. Clairic came in at 358 ms median and 736 ms at p95 on the same set. Cloudflare says responses stay under one second across its test workloads, well ahead of typical LLM response times.

Pricing and availability

The 9-billion-parameter Clairic-flash is priced at $0.10 per million input tokens, while the 27-billion-parameter Clairic will cost $0.24 per million input tokens. Until December 1, 2026, a promotional rate puts both models at $0.10, letting developers try the larger variant at the smaller one's price before the discount ends. Cloudflare bills only for input tokens and absorbs the cost of generated output, in keeping with what it calls its System One pricing principles.

The 9B model that previously served the preview @cf/cloudflare/clairic endpoint has moved to the flash endpoint, and the 27B model now serves the original one. Applications already pointed at the Clairic endpoint are upgraded in place, and the API schemas remain unchanged from the preview.

Why it matters

Extraction — pulling structured values out of receipts, leases, tickets and PDFs — is a workload where general-purpose LLMs tend to be slow, costly and imprecise. Cloudflare's argument, reinforced by rivals like Liquid AI and ScaleDown shipping their own extraction models, is that small, purpose-built models can handle these tasks at a fraction of the cost and latency. Publishing the weights invites the kind of experimentation and tuning that followed the open release of Clef, and sub-second response times matter for anyone embedding extraction into interactive products. For Cloudflare, the move also signals ambition beyond infrastructure: a frontier research team shipping and open-sourcing models of its own.

  • #cloudflare
  • #open-weights
  • #ai-models
  • #document-extraction
  • #machine-learning

Related posts