TL;DR
- Meta’s strategy shift: After positioning itself as the open-weight AI champion with Llama, Meta is building proprietary frontier models
- Mango: Vision/image generation model, quietly released as “Muse Image” — codename confirmed
- Avocado: Frontier LLM for text and code, delayed from March to at least May 2026 after internal tests showed performance between Claude Opus 3.5 and 4.0
- $14.3B Scale AI deal: Meta acquired a 49% stake in Scale AI, the largest single talent+data acquisition in AI history
- Impact: The industry’s most influential open-weight advocate is closing the door, leaving Moonshot and DeepSeek as the primary open-weight flag bearers
- Llama not dead: Meta continues Llama releases, but “frontier” investment is shifting to proprietary models
The Timeline of a Pivot
Meta’s relationship with open-source AI follows a clear arc. Understanding where it’s going requires understanding where it came from.
| Phase | Period | What Meta Did | Signal |
|---|---|---|---|
| Open champion | 2023-2024 | Released Llama, Llama 2, Llama 3 — open-weight, commercially usable | “Open-source AI is the path forward” |
| Market skepticism | Early 2025 | Llama 3.1, 3.2, 3.3 — still open, but smaller models | Questions about frontier commitment |
| The pivot | Dec 2025 | CNBC reports Meta building Avocado, a proprietary frontier model | CNBC exclusive |
| Mango confirmed | Early 2026 | Reports of Mango, a proprietary vision/image generation model | “Muse Image” released, codename Mango |
| $14.3B Scale AI | April 2026 | Meta buys 49% of Scale AI — largest AI talent acquisition ever | CNBC |
| Avocado delayed | March-May 2026 | Internal tests show Avocado between Opus 3.5 and 4.0 — not frontier enough | |
| Llama 4 | April 2025 | Meta’s last major open-weight release | Open, but not frontier |
(Sources: CNBC, WinBuzzer, TechBloat)
The Two Proprietary Models
Avocado (Frontier LLM)
Avocado was supposed to be Meta’s answer to GPT-5 and Claude Opus. It was designed as a proprietary frontier model for text and code — Meta’s first serious attempt to compete at the very top of the benchmark charts without releasing the weights.
The delay tells the story. Internal tests reportedly showed Avocado performing somewhere between Claude Opus 3.5 and 4.0 — solid, but not frontier-class in a world where Opus 4.8, GPT-5.6 Sol, and Kimi K3 are the benchmarks. Meta pushed the release from March to at least May 2026, and as of July, Avocado has not shipped. (Source: LinkedIn — Meta’s Avocado LLM Delayed)
Mango (Vision/Image Generation)
Mango shipped — but quietly. It was released as Muse Image, Meta’s latest image generation model, with the Mango codename confirmed internally. Unlike Llama, Muse Image is proprietary: no weights, no community fine-tuning, no open license.
Meta’s Instagram post from July 2026 confirmed the connection: “Just two days earlier, Meta quietly released something else. A new AI image generator called Muse Image, secretly codenamed Mango during development.” (Source: Instagram)
The $14.3B Question: Why Scale AI?
Meta’s 49% stake in Scale AI for $14.3 billion is the single largest AI talent and data acquisition ever. The rationale is clear: if Meta cannot compete on frontier model architecture alone, it will compete on data infrastructure.
Scale AI gives Meta:
- Access to the highest-quality human annotation pipeline in the industry
- RLHF and preference data at a scale no other lab (except perhaps OpenAI) can match
- A strategic blocker: Scale AI cannot work as closely with Meta’s competitors
The Scale AI deal is the most expensive acknowledgment yet that data quality, not model architecture, is the moat in frontier AI. (Source: royfactory.net)
Why This Matters
Meta’s pivot is consequential for three reasons:
1. The open-weight vacuum
Meta’s Llama releases were the primary reason enterprise teams believed in open-weight AI. If Meta is no longer investing its frontier R&D budget in open models, the open-weight mantle passes to DeepSeek (China, MIT license), Moonshot/Kimi K3 (China, Apache 2.0), and Mistral (France, Apache 2.0). For Western enterprises with data sovereignty concerns, the options narrow.
2. The competitive landscape without Llama
Llama’s open weights allowed thousands of companies to fine-tune, distill, and deploy models without API dependency. If Meta’s next frontier models are API-only, those companies face a choice: switch to API-dependent relationships with Meta, or adopt Chinese open-weight models — neither ideal for different reasons.
3. The paradox of open AI in 2026
Meta’s retreat is happening at the same time that Moonshot is releasing the largest open model ever (Kimi K3, 2.8T, Apache 2.0) and DeepSeek continues its MIT-licensed releases. The open-weight frontier is more alive than ever — but it is increasingly driven by Chinese labs, with Western champions retreating to proprietary models.
What’s Next for Meta AI
Meta’s strategy appears to be: keep Llama alive as a mid-range open offering, compete at the frontier with proprietary models (when Avocado eventually ships), and use the Scale AI data pipeline as the differentiator. The question is whether a “non-frontier open + frontier closed” strategy can sustain developer mindshare.
If Avocado never ships at frontier quality, Meta will have abandoned open-weight leadership for nothing. If it ships strong, it will face the same adoption barrier as every other proprietary API: why choose Meta’s API over Anthropic’s or OpenAI’s?
FAQ
Is Meta abandoning open-source AI? — Not entirely. Llama continues as a mid-range open offering. But frontier-level investment is shifting to proprietary models (Mango, Avocado).
What is Avocado? — Meta’s unreleased proprietary frontier LLM for text and code. Delayed from March to at least May 2026 due to sub-frontier benchmark performance.
What is Mango? — Meta’s proprietary image generation model, released quietly as “Muse Image.”
How much did Meta spend on Scale AI? — $14.3 billion for a 49% stake, the largest AI talent+data acquisition ever.
Who replaces Meta as open-weight champion? — DeepSeek (MIT), Moonshot/Kimi K3 (Apache 2.0), and Mistral (Apache 2.0) are now the primary open-weight frontier labs.