Canva's Magic Design pairs OpenAI's LLMs with its own design model

Two models doing two jobs

The most concrete engineering detail in this interview is how Canva builds Magic Design. Adams says the feature ties LLM prompting together with Canva's own design model to generate a design the user can then edit.

That is a specific division of responsibility, not a generic 'AI-powered' claim. The language model interprets the request; a separate, in-house design model produces the visual artifact. Canva names OpenAI and Leonardo.Ai as the outside collaborators, and states that Magic Studio is built partly on OpenAI's APIs.

For teams integrating foundation models, this is the recognizable pattern: use the general-purpose LLM for what it's good at, and keep a domain model for the output your product actually ships.

The 'last mile' stays manual

Adams repeatedly frames generation as a starting point, not an endpoint. He describes generating 'entire, editable designs,' and calls user editing 'the critical last mile.'

The stated strategy is 'the best of both worlds' — a generative experience for designs, images and text, plus a full workflow to manually edit each generated part, collaborate, and publish without leaving Canva.

This matters because it defines where the AI stops. The model gets a user to a polished draft faster; the human retains pixel-level control. Canva is not selling autonomous output, it's selling a faster entry into an editor it already owns.

Three sources of AI, and 225 million users to distribute it

Adams describes three ways Canva expands what AI can do on the platform: its own design AI research, collaborations with companies like OpenAI and Leonardo.Ai, and an ecosystem marketplace.

He cites 225 million active users and a developer marketplace where AI-powered apps — from AI presenters to music generators — plug directly into designs. That distribution base is what makes the ecosystem play viable: third-party AI tools reach users inside the product rather than as separate destinations.

With Magic Design, we tie together LLM prompting with our own design model to generate a unique design that our users can then edit for the critical last mile.Montana Labs

Internal adoption is tool-specific, not blanket

Canva's internal 'AI Everywhere' program is described in concrete, team-by-team terms: engineering uses Cursor for coding, finance uses a different set of tools, and company-wide hackathons let teams experiment outside normal workflows. The company also runs on ChatGPT Enterprise.

The framing — helping teams 'spot the right tools for their needs' — is worth noting. It's a portfolio approach to internal AI rather than a single mandated assistant.

What Canva's split architecture signals

The specific implication of this announcement is that a company with its own design model is still choosing to route prompting through OpenAI's APIs rather than build everything in-house. Canva keeps the model that differentiates its output — the design generation — and rents the general reasoning layer.

For applied teams weighing build-versus-buy, Canva's Magic Design is a usable template: own the model tied to your core artifact and your editing surface, integrate external LLMs for the interpretation layer, and design the handoff so users can still edit what the system produces.