Meta's Next Two AI Models Are Codenamed Avocado and Mango — Both Will Have Open-Source Versions
Meta's new AI chief Alexandr Wang confirmed open-source variants of upcoming models Avocado (LLM) and Mango (multimodal/video). The move reverses a December 2025 report that Meta was going fully closed-source.
Meta is back on open source — with caveats. The company’s new AI chief Alexandr Wang confirmed in a report published April 6-7, 2026, that both of Meta’s next major AI models will receive open-source releases. The models are internally codenamed Avocado (a large language model) and Mango (a multimedia and video generator).
This reverses — or at minimum significantly softens — a December 2025 Bloomberg report that said Meta was shifting entirely to closed-source AI development.
Who Alexandr Wang is and why he’s calling the shots
Wang joined Meta after its $15 billion acquisition of Scale AI in 2025, taking over as Meta’s AI chief. Scale AI built the data labeling infrastructure that powered training pipelines at every major AI lab. Wang understands what proprietary training data actually means — and that context shapes his thinking on the open/closed question.
His stated rationale for keeping open-source variants: he wants a U.S.-made, open alternative to Anthropic and OpenAI, which he sees as pivoting hard toward enterprise and government contracts. The argument is that open models are a strategic counterweight, not just a community goodwill play.
The Llama 4 problem
The background to this announcement is Llama 4 Maverick. At 400 billion parameters, it was supposed to be Meta’s competitive flagship — and it fell “significantly behind” frontier rivals. The gap between Meta’s open releases and the closed models from OpenAI and Anthropic widened noticeably in 2025.
Avocado and Mango are meant to close that gap. The closed proprietary versions will come first; open-source releases follow “eventually.” Wang didn’t give specific timelines.
What “open-source with caveats” means
The open-source variants will not be identical to the proprietary versions. Specific features will be stripped: likely reduced parameter counts in the public release, certain post-training steps (RLHF fine-tuning for specific capabilities), and possibly restricted code-generation capabilities tied to cybersecurity use cases.
This is a familiar pattern. Meta’s approach mirrors what the broader open-source AI community has been navigating for two years: “open weights” models that are technically downloadable but not identical to the internal version. The gap between open weights and true open source — including training data, full fine-tuning scripts, and complete system prompts — remains significant.
Why this matters for developers
Llama models run on a huge percentage of self-hosted AI deployments. Enterprises building internal tools, researchers fine-tuning for specific domains, and developers who won’t send data to external APIs all depend on capable open-weight releases from Meta.
If Avocado significantly outperforms Llama 4 Maverick and the open variant is competitive — even without the full feature set of the proprietary version — it refreshes Meta’s position as the default open alternative.
Wang’s hybrid strategy (closed first, open eventually) is also a funding and margin argument. Meta needs the proprietary versions to justify the capital expenditure to investors. Open source is the distribution flywheel that keeps the ecosystem locked to Meta’s architecture.
Neither Avocado nor Mango has a release date. Watch for the proprietary launches first; the open-source drop usually follows within two to six months.