AI provider / Meta
Meta
A provider long linked to open-weight influence and broad developer adoption, now mid-pivot to closed frontier models built by its new Superintelligence Labs unit.
Meta shaped the open-model era. LLaMA leaked within a week of its research release in 2023, and Llama 2 made open weights official and commercial — handing startups, researchers and enterprises a foundation they could build on freely.
From the 405B Llama 3.1, the first open frontier-class model, to multimodal and edge variants and the mixture-of-experts Llama 4, the line keeps pushing what open ecosystems can do. Its influence is measured less in a single product than in everything built on top of it.
In April 2026, that story split in two. The new Superintelligence Labs unit launched Muse Spark as a closed model, replacing Llama inside Meta AI itself — Llama 4 Behemoth never shipped, and by September Muse Spark 1.3 was Meta’s frontier model for long-horizon agentic work. Meta has said future releases will return to open weights, but for now its flagship assistant runs on a model nobody outside the company can download.
Model catalog
Every release in this provider’s model line.
Each row is a distinct model release — its date, its type and the traits it is known for. Open a source to dig into any of them.
| Model | Released | Type | Key traits | Source |
|---|---|---|---|---|
| Muse Spark 1.3 | Sep 2026 | Flagship |
| |
| Muse Spark | Apr 2026 | Closed release |
| |
| Llama 4 | Apr 2025 | MoE |
| |
| Llama 3.2 | Sep 2024 | Multimodal |
| |
| Llama 3.1 405B | Jul 2024 | Frontier open |
| |
| Llama 3 | Apr 2024 | Version update |
| |
| Code Llama | Aug 2023 | Coding |
| |
| Llama 2 | Jul 2023 | Open release |
| |
| LLaMA | Feb 2023 | Debut |
|
Timeline
A practical timeline of model releases and company milestones.
This page tracks the moments that matter most: major model releases, version jumps and company moves that changed how this provider shows up in the AI market.
Muse Spark 1.3 becomes Meta’s frontier model for long-horizon agentic work.
Now at a 1M-token context window and tuned for extended coding and agent sessions, it marks the clearest sign yet that Meta’s AI strategy now runs through closed, Superintelligence Labs-built models rather than open Llama releases.
Meta breaks from open weights, launching the closed Muse Spark.
Built over nine months by the new Superintelligence Labs, Muse Spark replaces Llama inside Meta AI with a natively multimodal model and a parallel-reasoning "Contemplating" mode — Behemoth never ships, and the planned follow-ups stay closed too.
Llama 4 moves the herd to mixture-of-experts.
Scout and Maverick debut natively multimodal MoE designs with a 10M-token context claim, amid a noisier reception than 3.1.
Llama 3.2 adds vision and pocket-sized variants.
11B and 90B vision models plus 1B and 3B for phones stretch the family from datacenter to device; 3.3 follows in December.
Llama 3.1 405B becomes the first open frontier-class model.
For one summer, the gap between open and closed frontier models effectively closes, validating Meta’s entire strategy.
Llama 3 brings 8B and 70B trained on 15T tokens.
A massive data-scaling bet pays off, with the 70B briefly the strongest open model and Meta AI launching across its apps.
Code Llama specializes the line for programming.
Built on Llama 2 with Python and instruct variants, it becomes the base of countless open coding assistants.
Llama 2 makes open weights official and commercial.
Free for most commercial use and distributed with Microsoft, it turns the leak-born ecosystem into Meta’s deliberate strategy.
LLaMA releases to researchers — and leaks within a week.
The weights escaping to the open internet accidentally founds the modern open-model ecosystem.