Technology releaseNovember 22, 2026Редакция 105.kz
Cloudflare releases Clairic and Clairic-flash weights: extracting data from PDFs and scans
Cloudflare has published open weights for Clairic and Clairic-flash — models for extracting structured fields from images and PDFs. We look at what this means for digitising equipment passports, reports and permits-to-work at enterprises in Kazakhstan, and what to check before a pilot.
Oil & GasEnergyUranium & MiningManufacturingPrivate LLMPredictive Maintenance
Key takeaway. Cloudflare has released the weights of the Clairic (27 billion parameters) and Clairic-flash (9 billion) models, designed to extract structured fields from images and PDFs. For Kazakhstani enterprises this is a component potentially suitable for on-premise deployment for digitising equipment passports, reports and permits-to-work. It must be verified independently: Russian and Kazakh languages, industrial documents and licence terms are not described in the source.
What happened
On 18 November 2026 Cloudflare published a post about two open-weight model releases: an updated Clairic and a new Clairic-flash. According to the company's description, Clairic-flash is a lighter and faster version of Clairic in a compact 9-billion-parameter build; the updated Clairic is a 27-billion-parameter model. Both are specialised in extracting data from unstructured sources: the full weights are published on Hugging Face, and cloud endpoints are available on Workers AI as @cf/cloudflare/clairic and @cf/cloudflare/clairic-flash. The same team, by its own statement, is now called the Workers AI superintelligence team and is responsible for Clef, Clairic and subsequent models.
The main change relative to the preview is a new architecture and work with PDFs. The company states that the model reads not only images but also PDF pages (up to four images or pages per request), and returns tabular line items row by row, preserving the "description — quantity — price" link. For text and numeric fields, instead of a single forward pass, parallel slot-wise autoregressive generation is used: each field stops at its own stop marker, and all fields advance simultaneously in one batch. Separate model heads have been added for calculations, estimates, forecasts, durations, computed dates and explicit arithmetic. The typed programmatic interface is retained: every response contains a confidence estimate, character offsets in the source text and probabilities. The stated latency is under one second; in the published measurements p50/p95 for text fields are 358/736 ms for Clairic and 155/328 ms for Clairic-flash. Pricing is $0.24 per million input tokens for Clairic and $0.10 for Clairic-flash, billed on input tokens only; until 1 December 2026 both models cost $0.10.
What this means for enterprises in Kazakhstan
The class of tasks such models cover is well known in industry: turning a scan or PDF into structured fields. For oil and gas these are passports of pump and compressor units, defect reports, operator round sheets, permits-to-work, laboratory test reports — data that today is entered into the maintenance system by hand. The model does not replace SCADA and does not read telemetry: it works with documents, and the result is then loaded into the maintenance system or into a digital operator advisor. The chain "historical failure event — cause — what was repaired" is reconstructed from reports and logs, and this is exactly the data layer that is usually missing for predictive maintenance (see /industry/solutions/predictive-maintenance).
In power generation the relevant documents are transformer test reports, switching schedules, and reports on feeders and ПСВ. In uranium and mining — commissioning test logs, measurement reports, repair documentation for mills and conveyors. At factories — invoices, waybills, batch certificates, incoming inspection reports.
Three environment constraints matter more than the model's capabilities. First: Russian and Kazakh are not mentioned in the source, so quality on your own corpus must be measured, not assumed. Second: open weights technically allow the model to be deployed on-premise, which matches the information security requirements of national companies and the logic of personal data localisation, but the licence terms are not described in the source, and this is checked before a pilot. Third: legacy process control systems (АСУ ТП) do not expose data directly — exchange with maintenance, ERP and the electronic archive usually goes through exports and an intermediate layer (estimate).
How to apply: an 8–12 week pilot
- Document audit. Select 3–5 document types and 300–500 instances (estimate) where manual entry is already measured in hours. Record the list of fields and the reference markup.
- Define the acceptance metric before the start. Not "the model works", but the share of fields accepted without correction, time per document and processing cost. The threshold is agreed with the process owner.
- Test the model on your own corpus on-premise: Russian and Kazakh text, low-quality scans, tables, dates. This is also where the confidence threshold and the manual review queue for fields below the threshold are configured.
- Integrate into the process. Extracted fields go into the maintenance system or archive, disputed ones go to manual review; permits-to-work and financial documents must not pass without a human.
- Measure the effect against the baseline: entry hours, error share, time to prepare periodic reporting before and after.
In a corporate environment such components are considered within the /industry/solutions/private-llm area — as part of a system, not as a standalone solution.
Limitations and risks
There are no public benchmarks for data extraction, which the company states directly: all figures are internal measurements, and the comparison results with D1.2 (combined with OCR) and Extract from ScaleDown the company itself calls indicative. There are no industrial examples in the source: it lists receipts, support tickets, lease agreements, credit pricing, traffic forecasting, expense auditing. Nothing is said about hardware requirements, the licence of the open weights, quality in non-English languages, or behaviour on poor scans.
Beyond that begins the customer's area of responsibility. The confidence estimate and per-field probabilities are not a guarantee: amounts, dates and numbers need a threshold and sample checks. The model extracts what is written and does not verify the physics of the process. Finally, data extraction is a component, not a system: without integration with the maintenance system, equipment master data and an acceptance procedure there will be no measurable effect.
Sources
- An updated Clairic and Clairic-flash are now open-weight, from the new Workers AI Superintelligence team - Cloudflare Blog · 2026-11-18T22:05:20.000Z
Both extraction models are hosted on Workers AI and full open weights are available on Hugging Face.
- Фото: Luc Viatour — Solar eclipse 1999 4.jpg (CC BY-SA 3.0, Wikimedia Commons)
This material was prepared with AI from the sources listed above. Spotted an inaccuracy? Write to info@105.kz.