Three New Models Announced

Meta has released three new Llama 4 models, each designed for different computational scales and use cases.

Llama 4 Scout features 17 billion active parameters with 16 experts and is the best multimodal model in its class. It offers an industry-leading context window of 10 million tokens and is pretrained and post-trained with 256K context length. Llama 4 Scout fits on a single H100 GPU with Int4 quantization and delivers better results than Gemma 3, Gemini 2.0 Flash-Lite, and Mistral 3.1 across a broad range of benchmarks. An experimental chat version scores an ELO of 1417 on LMArena.

Llama 4 Maverick has 17 billion active parameters with 128 experts and 400 billion total parameters. It beats GPT-4o and Gemini 2.0 Flash across a broad range of benchmarks and achieves comparable results to DeepSeek v3 on reasoning and coding at less than half the active parameters. The model fits on a single H100 host.

Llama 4 Behemoth contains 288 billion active parameters with 16 experts and nearly two trillion total parameters. It outperforms GPT-4.5, Claude Sonnet 3.7, and Gemini 2.0 Pro on STEM-focused benchmarks such as MATH-500 and GPQA Diamond.

Expanded Training and Multilingual Support

Llama 4 pre-training used over 30 trillion tokens, more than double the Llama 3 pre-training mixture. The models enable pre-training on 200 languages, including over 100 with over 1 billion tokens each.

Reduced Refusal Rates

Llama 4 refuses less on debated political and social topics overall, declining from 7% in Llama 3.3 to below 2%. The proportion of unequal response refusals is less than 1% on debated topical questions.

Availability

Llama 4 Scout and Llama 4 Maverick are available for download on llama.com and Hugging Face.

Source: Meta AI