In April 2025, Meta previewed a two-trillion-parameter model called Llama 4 Behemoth and told the world it would be the most capable open model ever released. It never shipped. Meta finished training it, judged the output quality too weak to publish, and quietly shelved it. Fourteen months later the company that made open-weight AI a mainstream idea has a closed frontier model, a superintelligence lab run by a former data-labeling founder, and no chief AI scientist. The open-source story everyone told about Meta in 2024 is over. Open weights are still reshaping the market. Meta just stopped being the company doing the reshaping.
That reversal is the real Llama story in 2026, and it changes how any buyer should think about betting infrastructure on Meta’s models.
Two things called “Meta AI,” and only one of them slipped
The naming confusion from 2024 is still worth clearing up, because the two halves of Meta’s AI effort moved in opposite directions.
The first is Meta AI the assistant: the chatbot inside WhatsApp, Instagram, Facebook, and Messenger. By 2026 Meta reports roughly a billion monthly users for it, up from about 500 million in 2024. Almost none of that growth came from people choosing Meta AI. It showed up in apps they already opened every day, and it expanded across 40-plus European countries after a March 2025 rollout. As a distribution story it worked. As a signal of model leadership it means very little, because the assistant runs on whatever model Meta points it at.
The second is Llama, the open-weight model family that developers download, fine-tune, and self-host. This is the part that mattered for the industry, and this is the part that lost momentum. The assistant is a product riding Meta’s install base. The models are infrastructure, and the infrastructure story turned in 2025.
What actually happened to Llama in 2025 and 2026
Llama 4 arrived in early 2025 in two configurations, Scout and Maverick, built on a mixture-of-experts design. Scout shipped with a headline ten-million-token context window, longer than anything commercially available at the time. The benchmarks underneath the headline were less impressive, and developers who tested Scout found the long-context claim more useful on a slide than in production, where retrieval-augmented pipelines still beat raw context stuffing for most work.
Then the top of the lineup fell apart. Behemoth, the roughly two-trillion-parameter flagship, was previewed in April 2025, then delayed over what SiliconANGLE reported as internal performance concerns. It was never released as an open model. Around the same delay, Meta reorganized. It founded Meta Superintelligence Labs on June 30, 2025, and it paid to staff it. Meta put $14.3 billion into Scale AI for a 49 percent non-voting stake, valuing the data company above $29 billion, and brought founder Alexandr Wang across to lead the new lab.
The philosophical break came next. Yann LeCun, the Turing Award winner who had been Meta’s chief AI scientist and the most articulate public voice for its open-science approach, left the company in November 2025. He went on to raise a reported $1.03 billion seed round for a Paris-based world-models startup, AMI Labs, and Meta declined to invest. When the person who spent a decade making your open-source argument leaves to build the opposite of a language model and you will not fund him, that is a strategy statement, not a personnel note.
The pivot finished in April 2026, when Meta Superintelligence Labs shipped Muse Spark, a closed-weight frontier model. We covered that shift when it happened in Meta’s $14 billion bet on closed-source AI. Read as a single arc, the sequence is unambiguous: preview a giant open model, fail to ship it, buy a data empire, lose the open-source evangelist, release a closed model instead. Llama did not get a proper successor. Meta chose a different lane.
“Open source” was always the wrong phrase for Llama
The 2024 framing treated Llama as open source. It never was, and that gap matters more now that the open-weight field has real alternatives.
Llama ships under Meta’s own Community License, not Apache 2.0 or MIT. As the Open Source Initiative has repeatedly pointed out, that license fails the open-source definition on three counts that no OSI-approved license contains. It sets a 700-million-monthly-active-user threshold, above which the free grant simply expires, a clause written to deny the largest competitors free use. It restricts entire categories of use. And it bans using Llama outputs to train other models, which is exactly the kind of downstream freedom an open license is supposed to protect.
For most small teams none of that ever bit. If you are a fifty-person company self-hosting Llama on your own GPUs, the 700-million-user ceiling is irrelevant. But “free to download and good enough” is not the same as “open,” and the distinction stopped being academic once Chinese labs started releasing genuinely permissive models that beat Llama on the benchmarks buyers actually care about.
The open-weight lead moved to China
Here is the part of the market that grew while Meta’s stalled. Through the first half of 2026, a run of open-weight releases from Chinese labs pushed real coding and reasoning scores to within a few points of the closed frontier, at a fraction of the price. On the current open-weight leaderboards, four of the five leading models come from Chinese labs, not from Meta.
DeepSeek’s V4 line ships under an actual MIT license and tops open leaderboards on agentic coding, a point we dug into when DeepSeek rebuilt its cheap model with post-training alone. Moonshot’s Kimi K3 became the largest open model ever released and posted the strongest scores among open weights on knowledge-heavy reasoning. Zhipu’s GLM family reached the frontier on domestic silicon, which we covered in the GLM-5.1 release built without Nvidia chips. Alibaba’s Qwen rounds out the top tier. Llama 4 Maverick is still a real model with a real Western deployment base, but it is now one entry on that list rather than the entry, and it competes under a more restrictive license than the models above it.
The irony is precise. The market Meta opened is bigger and more competitive than ever. The open-weight ecosystem is not shrinking; it is thriving. Meta simply is not the one leading it, and the leaders are shipping under the permissive licenses Meta never used.
Where Llama still earns a slot, from a practitioner’s chair
I evaluate self-hostable models for the same reason I always have: some data cannot leave the building, and a call to a hyperscaler API is a dependency I have to justify to a risk officer. From that seat, Llama is not dead, and pretending otherwise would be as lazy as the 2024 hype was.
Llama’s Western tooling gravity is real. The ecosystem that grew around it, Ollama for local runs, llama.cpp and GGUF quantization, the LoRA and QLoRA fine-tuning stack, still makes it one of the smoothest models to stand up on your own hardware. For an on-prem code assistant that never phones home, a mid-size Llama variant on a single accelerator is a defensible default. For regulated verticals fine-tuning on private clinical or legal corpora, the mature toolchain and Western provenance still count for something in a procurement review, where a US-licensed model is an easier conversation than a Chinese one regardless of benchmarks.
But the routing logic changed. In 2024 the honest recommendation was “reach for Llama first among open models.” In 2026 it is “qualify at least two open-weight models and let license, language of provenance, and cost-per-task decide.” If your compliance posture rules out Chinese models, Llama is your open answer almost by default. If it does not, DeepSeek or Qwen under MIT and Apache terms often win on both capability and licensing cleanliness. This is the same second-source discipline I apply to any critical vendor, and it is the practical takeaway from watching a single supplier go from category leader to one option among several. If you are standing this up yourself, our guide to self-hosting an open-source LLM in 2026 walks the stack, and the broader open-source versus closed AI decision frames when self-hosting is worth the operational cost at all.
The bet Meta actually made
The 2024 read on Meta was that open-sourcing Llama commoditized the model layer so Meta’s advantages, data and distribution and compute, could win above it. That logic still holds. Meta just concluded that giving away frontier weights was no longer the way to execute it. It kept the distribution engine, the billion-user assistant, and moved its best research behind a closed door, staffed by an acquired data operation and a superintelligence lab with a different founder’s fingerprints on it.
For a country building sovereign AI infrastructure, or a company deciding which weights to host, the lesson is not that Meta failed. Its assistant reaches more people than ever, and Muse Spark is a serious model. The lesson is that no single vendor owns the open-weight future, and the one that popularized it walked away from leading it inside eighteen months. Compute and control, not generosity, are what these companies optimize for, a pattern we traced in why every nation now treats compute like GDP. Bet your infrastructure accordingly: qualify a second model, read the license before the benchmark, and assume today’s leader is not guaranteed to be next year’s.