AI should trust the source in front of it.

Canonn builds context-faithful language models for work where the supplied document must outrank a model's memory. General-purpose models can confidently substitute a familiar policy, number, or clause for the one actually present in a customer's data. Canonn R1 is trained around a different contract: ground the answer in the provided material, extract exact details, decline when the answer is absent, and identify the supporting passage.

What we make

Canonn R1 is a 9B language model served through an OpenAI-compatible API. Teams use it for retrieval-augmented generation, customer support, policy search, contract review, and document question answering. The hosted service can retrieve across uploaded files. We publish benchmark methods and losing results alongside headline scores so teams can evaluate the product for the risks that matter to them.

How we work

Faithfulness is treated as a model capability, not a promise that prompt wording alone can guarantee. Product documentation describes limits, privacy behavior, pricing, and API semantics in public. Prompts and completions are not used for training; request-body logging is opt-in and off by default. Hosted inference runs in Canada. Canonn outputs still require human review for legal, medical, financial, safety-critical, or other consequential decisions.

For product, security, partnership, or research questions, visit the contact page. Engineers can start at the developer resource index, inspect the OpenAPI specification, or connect the live MCP server.