Meta releases Llama 4 Scout and Maverick

Background

The Llama family had been released as dense models, in which every parameter is used for every token. Llama 4 changed the design in two ways. It used a mixture-of-experts (MoE) architecture, which activates only part of the network for each token, and it was natively multimodal, meaning it was trained on text and images together rather than having vision added afterwards. Meta described Llama 4 as the start of a new era for these models.^[1]

What happened

On 5 April 2025 Meta released Llama 4 Scout and Llama 4 Maverick. They were the first Llama models built as mixture-of-experts and natively multimodal.^[1] In the same announcement Meta previewed a third model, Llama 4 Behemoth, which was still in training.^[1]

Key details

Scout. Scout has 17 billion active parameters and 109 billion total parameters. It has a context window of 10 million tokens.^[1] Active parameters are the ones used for each token. Total parameters include all the experts in the model.

Maverick. Maverick also has 17 billion active parameters, but 400 billion total parameters.^[1]

Behemoth. Meta previewed Behemoth with 288 billion active parameters and nearly 2 trillion total parameters. It was still in training at the time.^[1]

Training data. The models were pretrained on more than 30 trillion tokens covering 200 languages.^[1]

Leaderboard result. Meta said an experimental chat version of Maverick scored 1417 Elo on LMArena, a public ranking in which people compare model answers. LMArena later said Meta should have made clear that the version was a customised one, and it updated its policies after the episode.^[1][2]

Products. Meta AI in WhatsApp, Messenger and Instagram Direct, and on meta.ai, moved to Llama 4.^[1]

LlamaCon. At its first LlamaCon, on 29 April 2025, Meta announced a preview of the Llama API and said Llama had passed one billion downloads.^[3]

Delay of Behemoth. In May 2025 the Wall Street Journal reported that the release of Behemoth had been delayed.^[4]

Impact

Llama 4 showed Meta moving the Llama line to mixture-of-experts and natively multimodal designs, and it put the Llama models into Meta's consumer chat products. The LMArena episode led the ranking site to tighten its rules. Behemoth's delay and the later Meta reorganisation are covered in the entries that follow in this archive.^[1][2][4]