Cloudflare open-weights Clairic and Clairic-flash models
Cloudflare released Clairic and Clairic-flash as open-weight models, engineered by Workers AI Superintelligence for ultra-fast edge extraction and tool routing.
By Dillip Chowdary • Nov 19, 2026 • Source: Cloudflare Blog

At a glance
- Organization
- Cloudflare Workers AI Superintelligence Team
- Models Released
- Clairic and Clairic-flash
- License
- Open-weight under Apache 2.0
- Specialization
- Structured data extraction & tool routing
- Execution Target
- Cloudflare edge network & local hardware
- Inference Latency
- Sub-15ms for structured JSON outputs
In this post
Cloudflare's newly announced Workers AI Superintelligence team announced on Wednesday that an updated version of Clairic and its ultra-low latency companion Clairic-flash are now available as open-weight models. According to Cloudflare Blog's official release, the models have been published to Hugging Face under permissive open licenses, allowing developers to execute high-throughput structured entity extraction and agent dispatch on any hardware.
Rather than competing with trillion-parameter general-purpose conversational models, the Clairic architecture focuses explicitly on System One operational efficiency. By distilling extraction, schema enforcement, and tool classification into compact parameter footprints, Cloudflare delivers sub-second deterministic responses optimized for distributed edge execution.
01Specialized small models for the edge
Modern agentic architectures spend enormous token budgets on preliminary input parsing, classification, and schema conversion before ever invoking reasoning models. Running large frontier foundation models for trivial JSON schema conversion creates prohibitive latency bottlenecks and ballooning API bills across production pipelines.
Cloudflare engineered Clairic to solve this operational friction by optimizing specifically for dense document parsing, regex extraction, and deterministic tool call parameter formatting. The model processes unformatted web scrapes, PDF extracts, and API payloads directly into validated JSON schemas.
By trimming extraneous world knowledge and creative generation parameters, Clairic delivers specialized accuracy that matches or exceeds frontier models on strict schema-compliance benchmarks while running on commodity hardware.
02Clairic-flash architectural design
Clairic-flash represents the distilled, ultra-compact variant designed specifically for sub-15-millisecond Time-to-First-Token performance. It uses specialized attention heads optimized for key-value extraction, drastically reducing memory footprint during multi-tenant GPU concurrency.
During internal load testing across Cloudflare's global edge locations, Clairic-flash demonstrated near-instantaneous token generation when handling standard enterprise entity recognition workloads. The lightweight memory profile allows dozens of instances to reside concurrently on a single GPU node.
The model's compact geometry also makes it suitable for direct client-side deployment on developer laptops, mobile devices, and embedded edge gateways without requiring multi-gigabyte weight downloads.
Advertisement
Tech Pulse Daily
Get tomorrow's pulse first
Join engineers who read Tech Pulse before stand-up. Free, weekday mornings.
03Eliminating structured JSON parsing failures
A pervasive challenge in agent orchestration is output schema hallucination—where models omit required quotation marks, produce trailing commas, or inject conversational commentary alongside JSON objects. These syntax failures break automated downstream execution pipelines.
Clairic incorporates constrained grammar decoding natively into its fine-tuning objective, ensuring that outputs strictly adhere to provided Pydantic or TypeScript schema definitions without syntax deviations. The training dataset emphasizes hard edge cases including nested arrays, Unicode characters, and mixed-type dictionaries.
Benchmark evaluations demonstrate a 99.8 percent valid JSON compliance rate on first-attempt generation, completely eliminating the need for expensive retry loops or regex cleanup scripts.
04Direct deployment on Workers AI
Alongside the Hugging Face weights release, Cloudflare rolled out Clairic and Clairic-flash natively across its Workers AI serverless inference catalog. Developers can invoke both models with single-line bindings inside Cloudflare Workers, Pages, and serverless edge functions.
Deploying at Cloudflare's network edge places inference compute within milliseconds of global end-users, minimizing network transit latency for real-time applications such as form autocomplete, semantic routing, and security anomaly inspection.
Serverless billing for both models is structured on raw token throughput, providing cost reductions of up to 85 percent compared to standard frontier API endpoints.
05Open weights and local developer adoption
By releasing the model weights under an open-source license, Cloudflare enables machine learning engineers to self-host Clairic within private VPCs, air-gapped data centers, or local vLLM and Ollama instances. Full weight access also allows teams to conduct additional domain-specific fine-tuning.
Community developers immediately began experimenting with Clairic for local MCP tool dispatchers, automated git hook validation, and terminal-native log parsers. The release reinforces the growing industry momentum toward purpose-built small language models.
Cloudflare stated that the Superintelligence team plans to release additional open-weight artifacts over coming quarters, focusing on specialized reasoning verifiers and low-latency agent orchestration primitives.
Frequently asked questions
What are Cloudflare Clairic and Clairic-flash?
They are specialized open-weight language models designed for high-speed structured data extraction, schema compliance, and agent tool routing.
Where are the model weights available?
Cloudflare has published the weights on Hugging Face under the Apache 2.0 license for local or private hosting.
Can I run Clairic on Cloudflare Workers?
Yes, both models are natively available in the Cloudflare Workers AI catalog for instant serverless deployment.
How fast is Clairic-flash?
Clairic-flash achieves sub-15ms Time-to-First-Token latency, making it ideal for edge applications and real-time agents.
Sources
Author
Dillip Chowdary
Writes Tech Bytes coverage of AI, engineering, and the tools that actually ship. Editor of Tech Pulse Daily.
Related on Tech Bytes
Advertisement