Meta Plans to Open-Source New AI Models but Will Keep Its Most Powerful Closed

Meta is preparing to release its first AI models developed under chief AI officer Alexandr Wang, with plans to eventually offer open-source versions, Axios reported Monday. Before any public release, the company intends to keep some components proprietary and run safety evaluations, according to sources familiar with the plans. Not a full reversal. The announcement lands after months of speculation that Meta would abandon open source entirely, and it suggests the company has settled on a middle ground.

A hybrid bet, not a full retreat

Wang has indicated that some of Meta's largest new models will remain proprietary, with the company planning to release open-source versions of others over time. That is a sharp departure from the company's earlier approach, when it released full Llama model weights for anyone to download and modify at will.

The shift follows a bruising year inside Meta's AI divisions. DeepSeek's R1 model incorporated pieces of Llama's architecture to build a competitive system at a fraction of the cost. A Chinese rival, powered by Meta's own openness. People inside the AI division were livid. Llama 4 dropped in April 2025, and the response from developers was brutal. Coding benchmarks? Behind Google. Reasoning? Behind Anthropic and OpenAI. Not close, either.

But Wang has not abandoned the philosophy entirely. "We'll probably release a mixture of closed and open source models, but we're still very committed to open source as a company," he told the Economic Times in February. The difference now is operational. Meta picks what it shares and when, rather than defaulting to openness as a brand identity.

Avocado still isn't ripe

The new model family arrives under a cloud. Meta's next-generation model, codenamed "Avocado," was supposed to ship in March. It did not. Internal tests showed the model lagging behind Google's Gemini 3, OpenAI's GPT-5 updates, and Anthropic's Claude on reasoning and coding. The release slipped to May or June. Worse, Meta's leadership reportedly discussed licensing Gemini from Google, temporarily, to keep its own consumer AI products functional while engineers at the Menlo Park campus work through the problems.

Wang joined Meta through a $14.3 billion deal that gave the company a 49% stake in his startup Scale AI last June. He has since led Meta's Superintelligence Labs and the elite TBD Lab where Avocado is being built. The models now carry the full weight of that investment.

Meta acknowledges the gap with competitors. Sources told Axios the company "knows its new models may not be competitive across the board" with upcoming releases from rival labs, but believes it will find consumer-facing strengths others lack. The financial pressure behind that confidence is enormous. Meta raised its 2026 capital expenditure guidance to $115 billion to $135 billion. If you are spending at that scale, a model that merely "shows the rapid trajectory we're on," as a Meta spokesperson told CNET, does not reassure anxious investors.

Open source under pressure everywhere

Meta's hedging fits a pattern now visible across the industry. Alibaba recently kept its most powerful new Qwen models proprietary, reversing years of open releases. The playbook is consistent on both sides of the Pacific: seed the developer ecosystem with open models, then close the door on the systems that matter most. If you have watched the open-source AI space over the past year, this should feel familiar.

And yet open source still produces concrete results. Cursor, the AI coding startup valued at $29.3 billion, built its Composer 2 on top of Moonshot AI's open-source Kimi 2.5. Roughly a quarter of the final model's training compute came from that base. The rest was Cursor's own reinforcement learning, layered on top. One product launch, and the practical value of open weights snaps into focus. The tension between strategic caution and ecosystem demand is exactly what Meta is trying to thread.

Distribution as the real bet

Wang sees the competitive map differently from other lab leaders. OpenAI and Anthropic are increasingly focused on governments and enterprise contracts. Meta's play, per sources, targets consumers.

Three and a half billion people use Meta's platforms daily. AI already runs inside WhatsApp, Facebook, and Instagram. Free global services that no pure-play AI lab can replicate. Build the widest funnel, not necessarily the best model. That is the wager.

The competition is not waiting. Anthropic is testing a model it describes as a "step change" beyond its current offerings. OpenAI is signaling its own significant advances. Both could ship before Meta does.

What the first models under Wang actually look like, whether the open-source versions follow weeks or months behind the closed ones, and how Avocado performs against rivals that have already lapped it will reveal more about Meta's real AI strategy than anything its executives have promised so far.

Alibaba Releases Third Closed-Source AI Model in Three Days, Signals Profit-First Strategy

Three proprietary AI models in three days. That's what Alibaba shipped this week, with Qwen3.6-Plus landing Thursday as the third, Bloomberg reported. This one handles agentic coding and multimodal re

The Implicator

Nvidia's Open Source Play Isn't About Openness

OpenAI is welding together its own chips. Google has TPUs humming in data centers across three continents. Anthropic and Amazon are building custom silicon. The companies buying Nvidia's $40,000 H100s

The Implicator

America's AI Startups Are Building on Chinese Foundations

Martin Casado dropped a number that should unsettle Washington. The Andreessen Horowitz partner told The Economist that up to 80% of AI startup pitches his firm receives now feature American companies

The Implicator

Maria Garcia

Los Angeles

Tech culture and generative AI reporter covering the intersection of AI with digital culture, consumer behavior, and content creation platforms. Focusing on technology's beneficiaries and those left behind by AI adoption. Based in California.