Cloudflare gives extraction its own model
Not every step in your AI workflow needs a general-purpose model.
In today’s edition
- Cloudflare previews a model built for extraction.
- DeepSeek opens the inference engine behind Flash.
Cloudflare gives data extraction its own model
Cloudflare has previewed Clairic on Workers AI, a model that returns typed values from messy inputs without generating text. Its worked support example makes a concrete case for separating extraction and decisions from writing the reply.
- Requests combine text, up to four optional images, and named fields. Answers include source spans, confidence scores, or uncertainty ranges, depending on the field type.
- Cloudflare’s support-workflow test used Clairic for extraction, Clef for the decision, and GLM-5.3 for the reply. It finished in 19.3 seconds versus 61.7 seconds for an all-GLM-5.3 version, at 49% lower cost.
- Both versions got the facts and decision right in every reported run. Clairic’s repair estimate stayed identical, while GLM-5.3’s estimate and ranges varied.
- The hosted endpoint is an early checkpoint. Open weights are planned, not released, and Cloudflare says arithmetic, estimation, and forecasting need more work.
What it means Separate extraction, decisions, and writing when your workflow needs auditable values; Cloudflare’s example shows a concrete alternative to asking one generative model to handle everything.
Source: blog.cloudflare.com
DeepSeek opens the engine behind its Flash model
DeepSeek released DeepInfer on October 20, opening the inference framework behind its large-scale production services. Builders can now inspect and adapt the engine behind DeepSeek-V4.1-Flash, rather than work only through the hosted API.
- The repository uses Apache 2.0 licensing and provides offline inference. It is not a complete inference service.
- The release includes distributed inference, speculative decoding, optimized kernels, and structured outputs.
- DeepSeek also released deepseek-recipe, its production token encoding and decoding component, aligned with the processing logic behind the official API.
What it means Self-hosting teams gain reusable inference code and API-processing logic, but should budget for service infrastructure rather than treat this release as a turnkey deployment.
Source: deepseek_ai