Llama 4 Maverick

What it is

Llama 4 Maverick is a natively multimodal mixture-of-experts model, released on 5 April 2025 alongside Llama 4 Scout. It has 17 billion active parameters, 128 routed experts plus a shared expert, and 400 billion total parameters. It alternates dense and mixture-of-experts layers, and it fits on a single NVIDIA H100 DGX host.^[1][2] Because only some experts run for each token, the 17 billion active parameters are the figure that matters for compute, while the 400 billion total is the size of the full network.

Maverick was codistilled from Llama 4 Behemoth, the larger model Meta previewed but did not release. Codistillation means a smaller model is trained alongside a larger teacher model and learns from its outputs.^[1]

Origins and training

The Llama 4 herd was trained on over 30 trillion tokens across 200 languages, more than double the Llama 3 training mixture. Post-training pruned about 50% of the supervised fine-tuning data that was judged easy.^[1]

Maverick itself was trained on about 22 trillion tokens with an August 2024 knowledge cutoff. Training used 2.38 million GPU hours.^[2]

Main features

Context window. Maverick has a context window of 1 million tokens.^[2]

Benchmark results. On instruction-tuned benchmarks, Maverick scored 80.5 on MMLU Pro, 69.8 on GPQA Diamond, 43.4 on LiveCodeBench, 73.4 on MMMU, 90.0 on ChartQA, 94.4 on DocVQA and 73.7 on MathVista.^[2]

Comparisons Meta made. Meta said Maverick beat GPT-4o and Gemini 2.0 Flash on a broad range of benchmarks. It said Maverick was comparable to DeepSeek V3 on reasoning and coding while using less than half the active parameters.^[1]

Chat ranking. Meta said an experimental chat version of Maverick scored an Elo of 1417 on LMArena, a public leaderboard that ranks models by human votes.^[1] That entry, "Llama-4-Maverick-03-26-Experimental", was a version optimised for conversationality rather than the public model. LMArena said Meta should have made this clearer, and it published more than 2,000 battle results and updated its policies.^[3]

When the unmodified public release, Llama-4-Maverick-17B-128E-Instruct, was added to LMArena, it ranked 32nd.^[4]

Training mix. The herd's training data spanned more than 30 trillion tokens. Maverick is one member of that herd, so the same mix stands behind its training.^[1]

Licence and access. Maverick was released under the Llama 4 Community License Agreement, on llama.com and Hugging Face, and was offered in the Llama API preview. At LlamaCon on 29 April 2025, Meta reported that Llama had passed one billion downloads.^[2][5]

Its place in development

Meta's planned follow-up models, Llama 4.1 and 4.2, were mentioned in a June 2025 memo that created Meta Superintelligence Labs.^[6] Meta's next flagship released publicly was Muse Spark, in April 2026.^[6][7]

The Maverick case shows how a model's public ranking can depend on which version was tested. The published result came from a conversational variant, while the public model that developers downloaded ranked lower.^[3][4]