Sources: Meta's new AI model, codenamed Avocado, may launch in spring 2026 as a “closed” model, and was trained using Google's Gemma, OpenAI's gpt-oss, and Qwen
Meta Platforms Inc.'s Mark Zuckerberg, months into building one of the priciest teams in technology history …
Context & Ripple Effects
Avocado emerged in related coverage as Meta's planned successor to Llama and a frontier-model effort. This report adds two strategically important details: a possible proprietary release and reported training inputs from models associated with Google, OpenAI and Qwen.
Subsequent reporting positioned Avocado within a broader Meta model slate, including the image- and video-focused Mango model effort. Internal memos later described Avocado's claimed text-task efficiency gains over Llama 4 Maverick, making the release model consequential alongside raw capability.
First-order effects
- If Meta ships Avocado as reported, developers and enterprises would face a different access proposition from the Llama line: use of a proprietary Meta model rather than downloadable model weights.
- The reported use of Gemma, gpt-oss and Qwen as training inputs would make model provenance and the terms governing those inputs a more immediate issue for Meta's launch and deployment teams.
Second-order effects
- A closed Avocado would sharpen pressure on Meta to differentiate through performance, product integration and commercial terms, rather than relying primarily on open-weight distribution.
- Google, OpenAI and Qwen would gain another high-profile example of their models functioning as inputs to a rival's development process, increasing attention to how model providers define and enforce permitted downstream use.
Third-order effects
- If leading developers increasingly combine outside models during training while restricting their own frontier releases, the market could separate more clearly between open model availability and closed frontier commercialization.
- That split would make frontier-model access governance more central: the decisive questions become not only who can build models, but who controls weights, APIs, training inputs and downstream rights.
The trend: AI developers are converging on hybrid strategies that draw on open-model ecosystems during development while reserving their strongest models for controlled distribution.