Llama 4 is a family of mixture-of-experts (MoE) language models released by Meta AI on April 5, 2025, comprising Llama 4 Scout, Llama 4 Maverick, and the preview-only Llama 4 Behemoth. It was Meta's first MoE generation, following dense Llama predecessors and the MoE designs of DeepSeek-V3, Qwen3, and Kimi K2. Its reception was shaped by a benchmark-gaming controversy on LMSYS Chatbot Arena, by independent reproductions that fell short of Meta's reported scores, and by the non-release of Behemoth; commentary including (Source: epochai.substack.com) and State of AI Report 2025 read the release as the point at which Meta's open-weight commitment began to weaken.
| Field | Value | |
|---|---|---|
| Developer | [[meta | Meta AI]] |
| Released | April 5, 2025 | |
| Architecture | Mixture-of-Experts — Meta's first MoE generation | |
| Family | Llama 4 Scout (109B total / 17B active, 16 experts); Llama 4 Maverick (400B total / 17B active, 128 experts); Llama 4 Behemoth (~2T total / ~288B active — preview only, never released) | |
| Context length | 10M tokens (Scout); 1M (Maverick) — marketed; real-world long-context performance contested | |
| Training tokens | ~30T | |
| License | Llama 4 Community License (tightened from Llama 3) |
Training and architecture
Llama 4 marked Meta's transition from dense to MoE architectures. Scout has 109B total parameters with 17B active, using 16 experts with 1 active per token plus a shared expert, and was designed for single-H100 inference. Maverick has 400B total parameters with 17B active, using 128 experts with 1 active per token plus a shared expert, and interleaves dense and MoE layers. Behemoth was previewed as a ~2T-parameter model (~288B active) still in training in April 2025. The family was trained on roughly 30T tokens.
Two architectural features distinguish the family. iRoPE (interleaved Rotary Position Embeddings) alternates layers with and without positional encoding, supporting the 10M-token advertised context length for Scout (1M for Maverick). The models also use native multimodality, with early-fusion vision and text training across the family, in contrast to prior Llamas that added vision post-hoc (Llama 3.2-Vision).
Capabilities and benchmarks
Marketed context lengths were 10M tokens for Scout and 1M for Maverick, though real-world long-context performance was contested. Independent reproductions of Meta's claimed MMLU-Pro, GPQA, and coding numbers fell short of the reported scores. The gap was small but consistent enough that third-party leaderboards de-prioritized Meta's numbers.
Benchmark-gaming controversy
Within days of release, researchers found that the Llama 4 Maverick version submitted to LMSYS Chatbot Arena was an "experimental chat version" tuned for human-preference evaluation rather than the released weights. LMSYS subsequently clarified that the ranked model was not the public release. The episode was widely read as an attempt to game a public benchmark, and is cited as one of the more prominent cases of benchmark-gaming by a major lab (see AI Benchmarks).
License
The Llama 4 Community License retains the Llama 3 restrictions — a 700M monthly-active-user cap, attribution requirements, and a "no improving other LLMs" clause — and adds further terms. Companies headquartered in the EU are barred from using Llama 4 multimodal features, which Meta tied explicitly to EU AI Act GPAI obligations and framed as a consequence of regulatory uncertainty. The acceptable-use policy was expanded with additional restrictions on use in "critical infrastructure" and defense applications, with carve-outs requiring Meta approval. The tightening relative to Llama 3 fits a broader pattern in which Chinese labs have shipped more permissively licensed weights than US labs at the frontier (see Open-Source AI / Open-Weight Models).
Reception and role in Meta's open-weight arc
(Source: epochai.substack.com) and State of AI Report 2025 read Llama 4 as the point at which Meta's "open frontier" brand started to look contingent. The 2T-parameter Behemoth, previewed as still training in April 2025, was never publicly released. Its absence, combined with Meta's later hiring of Alexandr Wang, the Scale AI investment, and the restructuring of FAIR into a "superintelligence" unit, is read as a pivot toward closed frontier development. As of April 2026, Meta had not released a Llama 5 or a clearly open successor to Llama 4. (Source: epochai.substack.com)
In commentary on open-source AI, Llama 4 is treated as an example of a US "retreat from open"; in the framing of US-China AI competition, Meta's pullback positions Chinese labs as the dominant suppliers of open-weight frontier models. For Meta, the release marks a shift in corporate strategy from open-default toward closed development, a reading also associated with Mark Zuckerberg.
Related models
Llama 4 supersedes Llama 3 (Meta) and is contemporaneous with the MoE designs of DeepSeek-V3, Qwen3, and Kimi K2.
Relationships
- instance-of: General-Purpose AI (GPAI)
- supersedes: Llama 3 (Meta)
- related: Meta AI, Open-Source AI / Open-Weight Models, Qwen3, DeepSeek-V3, Kimi K2, Mark Zuckerberg, EU AI Act (Regulation 2024/1689)
- contradicts: Open-Source AI / Open-Weight Models narratives that frame Meta as a durable open-weight champion.