Meta has won over developers in recent years with its Llama series—Llama 2, Llama 3, Llama 4, each more powerful than the last, all with open weights and a thriving community. But reports from Axios and SiliconANGLE last week introduced a new twist: Meta's next two flagship models in development, Avocado and Mango, will not be fully open-sourced.
This development deserves a closer look.
What Are Avocado and Mango?
According to current reports, Meta Superintelligence Labs is running two projects simultaneously:
- Avocado: A pure text LLM focused on coding and reasoning, positioned against OpenAI's o-series and Anthropic's Claude.
- Mango: A multimodal image/video generation model, competing with Sora and Google Veo.
Both models are expected to launch internally at Meta in the first half of 2026, with external releases to follow. The effort is led by Alexandr Wang—yes, the co-founder of Scale AI. Meta acquired a 49% stake in Scale AI last year at a $14.3 billion valuation, bringing Wang in to head Meta's AI division.
Open-Source Versions Exist, But You Won't Get the Full Package
Here's a key detail: Meta is not saying it will completely close-source these models, but rather partially open-source them.
According to Axios, the open-source versions may lack the following:
- Reduced parameter count: The community version may have fewer parameters than the flagship model.
- Incomplete MoE architecture: Certain neural networks within the mixture-of-experts structure may not be included in the open-source release.
- Skipped post-training steps: This means outputs from safety alignment, RLHF, and similar processes may not be fully included.
The most critical restriction: Avocado is reportedly highly capable in cybersecurity code generation, but this capability will not appear in the open-source version. The stated reason is security concerns, to prevent misuse.
How Does This Compare to Llama 4?
Llama 4 was just released, with Maverick featuring a 400B-parameter MoE architecture and open weights that anyone can download and run. This remains Meta's fully open-source approach.
Avocado and Mango represent the next step. Meta has likely concluded that the fully open Llama series has already built a strong enough developer ecosystem and brand recognition. The next move is to create truly frontier models, and at this level, full open-sourcing would directly erode competitive advantage.
In other words: Llama is a gift to the open-source community; Avocado/Mango are Meta's own weapons.
Axios noted that Meta does not expect these new models to lead in every metric, but they will have advantages in several specific areas. This statement itself shows Meta knows the battle is tough and is choosing a few key directions to focus on.
What This Means for the Open-Source Community
One major reason Meta previously pushed full open-source for Llama was that it helped Meta resist ecosystem monopolization by Google and OpenAI. When developers use Llama, they stay in Meta's orbit. But when Meta itself wants a piece of the frontier model pie, the cost of full open-source becomes too high.
Partial open-source is a compromise, but for the open-source community, the line is delicate: you can still get a decent version, but not the most powerful one.
If Avocado truly outperforms Claude Opus or the GPT-5 series in coding ability, the missing part is not just a feature gap—it's a competitive gap.
The Bigger Picture
Don't forget the context: Meta announced in 2025 that it would invest $60-65 billion in AI for the year, with data center construction accounting for the bulk. Since Alexandr Wang took the helm, Meta AI's focus has shifted from generously giving away open-source to securing a place in the frontier model race.
Llama earned Meta developer trust. Now Meta wants to use that trust to place higher bets—in commercial deployment, in the AI assistant ecosystem, and in head-to-head competition with OpenAI and Anthropic.
Is partial open-source a temporary compromise in this transition, or a long-term strategy? We'll know after seeing the reaction to Avocado's release.
Sources: CocoLoop, Report: Meta developing open-source versions of upcoming AI models (SiliconANGLE), Meta is developing a new image and video model for a 2026 release, report says (TechCrunch), Meta readies next-generation Mango and Avocado AI models for 2026 launch (mlq.ai)