What Happened
Meta’s next-generation foundation model “Avocado” has experienced two delays:
- Original target: Late 2025
- First delay: March 2026
- Second delay: May 2026 or later
The reason for the delay is critical: internal testing found that while Avocado outperformed Llama 4 and older versions of Gemini, it still fell short of Meta’s expected “frontier” standards on certain core metrics.
More notably, the strategic shift accompanying the delay: Meta is moving from “full-force open-sourcing Llama” to a “dual-track open and closed source approach.”
Why Meta Is Shifting
1. The “Double-Edged Sword” Effect of Open Source
The success of the Llama series brought Meta enormous industry influence, but also created problems:
- Competitors riding free: Chinese model manufacturers (DeepSeek, Qwen, etc.) are rapidly iterating on the Llama architecture, in turn threatening Meta’s competitiveness
- Commercialization difficulties: Open-source models make it hard for Meta to monetize directly through model APIs
- Security and compliance pressure: Open-source models could be used for malicious purposes, exposing Meta to regulatory accountability
2. Internal Route Disputes
According to multiple media reports, there are internal divisions at Meta regarding the open-source approach:
- Zuckerberg: Insists on the open-source strategy, viewing it as Meta’s differentiating weapon against Google and OpenAI
- Some technical executives: Believe open source is harming Meta’s commercial interests, advocating for a shift to closed source
3. Capital Expenditure Pressure
Meta’s 2026 capex guidance reaches $125-145B (recently raised by $10B), primarily for AI infrastructure. With such massive investment, Meta needs to prove its AI investments can generate direct commercial returns — open-source models struggle to do this.
Meta’s New Three-Track Strategy
| Track | Content | Status |
|---|---|---|
| Closed-source frontier | Avocado model (proprietary, available only through Meta platform) | In development, expected May+ |
| Open-source follow-up | Llama 5 (open-source, but likely 1-2 generations behind Avocado) | Planned |
| Application integration | Deep AI integration into Facebook, Instagram, WhatsApp | Ongoing |
This strategy is increasingly similar to Google’s Gemini approach: keep the best models proprietary, release open-source versions that are “good enough but not the strongest.”
Impact on the Open-Source AI Ecosystem
Meta’s strategic shift is a blow to the open-source AI community:
- Llama’s “benchmark” status may waver: If Meta no longer open-sources its strongest capabilities, the Llama series may downgrade from “industry benchmark” to “reference implementation”
- Chinese model manufacturers benefit: Qwen, DeepSeek and others have already established strong competitiveness in the open-source space. If Meta reduces open-source investment, these manufacturers may fill the gap
- “Trust crisis” for the open-source community: If Meta’s open-source commitments begin to waver, other major companies (Google, Microsoft) may follow, and the open-source AI ecosystem could enter a “pseudo-open-source” era
Recommendations for Developers and Enterprises
| Your Role | Recommendation |
|---|---|
| Developers dependent on Llama | Evaluate alternatives: Qwen3.6 open-source series, GLM-5.1 open-source weights |
| Enterprise users | Don’t bet your technology roadmap on a single vendor’s open-source commitments; build multi-model routing capabilities |
| Open-source community contributors | Watch the community building around Chinese open-source models like Qwen and DeepSeek, which may become the core of the next-generation open-source ecosystem |
Landscape Assessment
Meta’s strategic shift reflects a broader industry trend: the “honeymoon period” of open-source AI is ending.
From 2023-2025, open source was the weapon to challenge closed-source giants. But by 2026, as the capability gap between closed and open-source models narrows and the commercialization pressure on open-source models increases, the question “why open-source?” is being taken more seriously by more people.
For Chinese AI companies, this is actually an opportunity window. If Meta and OpenAI are both reducing open-source investment, Chinese open-source models (Qwen, DeepSeek, GLM) may become the new leaders of the open-source AI ecosystem.