Meta Platforms released Llama 4 Scout and Llama 4 Maverick on April 5, 2025, calling them its most advanced models yet and the best in their class for multimodality. The two models are the first in the Llama 4 herd to ship publicly, and both are available for download on llama.com and Hugging Face.
Llama 4 Scout is a 17 billion active parameter model with 16 experts, built to fit on a single NVIDIA H100 GPU with Int4 quantization. It carries a 10 million token context window, which Meta says is industry-leading. Llama 4 Maverick, also 17 billion active parameters but with 128 experts, fits on a single H100 host and is positioned as the stronger of the two for image and text understanding.
The release matters now because Meta is spending up to $65 billion this year on AI infrastructure, and investors want to see returns. Open-weight models are Meta's counterweight to OpenAI's and Google's closed systems. If developers adopt Llama 4 widely, Meta gets influence over the stack without charging per token.
What Scout and Maverick Can Do
Both models are natively multimodal, meaning they process text and images in the same architecture rather than bolting a vision encoder onto a text model. Scout handles multilingual text and image input across 12 languages including Indonesian, Thai, and Vietnamese. Maverick adds tool-calling and agentic system support, which Meta says suits coding and multi-step tasks.
Meta measured Scout against Gemma 3, Gemini 2.0 Flash-Lite, and Mistral 3.1 on a range of benchmarks and reported Scout ahead in its class. Maverick scored an ELO of 1417 on LMArena in its experimental chat version and beat GPT-4o and Gemini 2.0 Flash across a broad set of benchmarks, while matching DeepSeek v3 on reasoning and coding at less than half the active parameters.
The Behemoth Teacher Behind Them
Meta also previewed Llama 4 Behemoth, a 288 billion active parameter model with 16 experts that it calls one of the smartest LLMs in the world. Behemoth is not shipping yet. It is still training, and Meta used it as a teacher model to distill knowledge into Scout and Maverick.
According to Meta's technical details, Behemoth outperforms GPT-4.5, Claude Sonnet 3.7, and Gemini 2.0 Pro on STEM benchmarks including MATH-500 and GPQA Diamond. That is a strong claim, and it is self-reported. No independent evaluation of Behemoth exists because the weights are not public.
The distillation story is the real bet. Meta is betting that a teacher model no one can run will produce student models everyone can. That only works if the distillation quality holds up under independent testing, which has not happened yet.
The Launch Was Not Smooth
The Information reported on Friday that Meta delayed the Llama 4 launch because the model did not meet expectations on technical benchmarks, particularly reasoning and math tasks. The company was also concerned that Llama 4 was less capable than OpenAI's models at humanlike voice conversations.
That reporting is worth weighing against the launch-day claims. Meta said the models are its best yet. The Information said they fell short of internal targets. Both can be true if the internal targets were set above what shipped, and that is the more likely reading.
Meta has a pattern of releasing models that trail the frontier on some benchmarks while leading on cost and openness. Llama 4 fits that pattern. Whether it closes the gap with closed competitors on reasoning is the open question.
What This Means for Developers
Scout's 10 million token context window is the headline feature for anyone building document-heavy applications. A 10 million token window can hold roughly 7,500 pages of text, which means a single prompt can ingest an entire codebase or a large legal filing without chunking.
Maverick's cost position is the other draw. Meta says it delivers comparable results to DeepSeek v3 at less than half the active parameters. Fewer active parameters means lower inference cost, which matters for startups that cannot afford frontier-model pricing.
Both models support 12 languages, including Indonesian. For developers in Southeast Asia, that is a practical consideration. Most frontier models handle English well and other languages unevenly. A model trained with Indonesian in its supported language list is more likely to produce usable output for local applications.
The Money Behind the Models
Meta plans to spend as much as $65 billion this year to expand its AI infrastructure, according to Reuters. That figure sits alongside investor pressure on big tech to show returns on AI spending. Meta's bet is that open models create an ecosystem that drives usage of its own products, from WhatsApp to Instagram to the Meta.AI website.
The open-source angle is not charity. Meta benefits when developers build on Llama because those applications run on Meta's infrastructure and feed data back into Meta's products. It is a platform strategy, not a giveaway.
Big technology firms have been investing aggressively in AI infrastructure since OpenAI's ChatGPT changed the market. Meta is the only one of the major players releasing its top models as open weights. That distinction is the strategic bet, and it is the reason Llama 4 matters beyond the benchmark numbers.
What to Watch Next
Behemoth needs to finish training and ship. If it does, and if the distillation quality holds, Meta will have a credible frontier model that no competitor can match on openness. If Behemoth stalls, Scout and Maverick are strong mid-tier models but not category leaders.
Independent benchmarks are the next checkpoint. Meta's claims on Scout and Maverick are self-reported. LMArena and other third-party evaluations will either confirm or complicate the picture in the coming weeks.
The developer adoption question is the one that decides whether this matters. Open weights only matter if people use them. Meta has the distribution through Hugging Face and its cloud partners. Whether that translates into production deployments is something the numbers will show by mid-2026.
For more on how open-weight models compare to closed systems, see Reuters coverage of AI. For more on this site's AI coverage, browse the archive.