Meta Releases Llama 4 Models With Multimodal Capabilities

What Happened

Meta released the Llama 4 family of open-weight large language models, marking the company's first publicly available models with native multimodal capabilities spanning text and image inputs. The release makes Llama 4 available to developers and researchers under Meta's custom community license.

The Models

The Llama 4 release includes multiple model variants at different scales. The family introduces a mixture-of-experts architecture, a design that activates only a subset of model parameters during any given inference step. Meta stated this approach improves computational efficiency relative to dense model designs of comparable parameter counts.

The models accept both text and image inputs, extending the capabilities of prior Llama generations, which were limited to text. Meta has described Llama 4 Scout and Llama 4 Maverick as the initial releases in the family, with a larger model referred to as Llama 4 Behemoth disclosed as still in training.

Llama 4 Maverick features 400 billion total parameters with 17 billion active parameters per forward pass. Llama 4 Scout features 109 billion total parameters with 17 billion active parameters per forward pass. Meta reported that Scout supports a context window of 10 million tokens, which the company described as among the longest available in a publicly released model.

Background

Meta has released successive generations of Llama models since 2023, positioning the open-weight series as an alternative to closed commercial models from OpenAI, Anthropic, and Google. The Llama 2 and Llama 3 families drew broad adoption among developers and enterprise users building on open-weight foundations.

The company has consistently argued that open model releases benefit the broader research community and enable independent safety evaluation. Meta AI Research leads development of the Llama series.

Prior to this release, multimodal capabilities in the Llama line were available only through separate model variants not integrated into the core Llama release. Llama 4 consolidates text and image understanding into a unified model architecture.

Benchmark Performance

Meta reported results across a range of standard evaluation benchmarks. The company stated that Llama 4 Maverick achieved scores competitive with GPT-4o and Google's Gemini 2.0 Flash on several academic benchmarks, including multi-step reasoning and image understanding tasks. Meta noted Maverick ranked first on the LMArena human preference leaderboard at the time of release, though leaderboard rankings shift as new models are submitted.

Independent verification of these results by third-party evaluators had not been published at the time of the announcement.

Availability and Licensing

The models are available for download through Meta's official Llama website and via Hugging Face. Meta's community license permits commercial use subject to usage restrictions, including a provision requiring companies with more than 700 million monthly active users to seek a separate license from Meta.

Meta also made Llama 4 available through its own AI assistant products and through cloud provider partners including Amazon Web Services, Google Cloud, and Microsoft Azure.

What It Means in Practice

Developers can use Llama 4 models to build applications that process both text and images without relying on closed commercial APIs. The mixture-of-experts architecture means the models can be served at lower per-token compute cost than dense models with equivalent total parameter counts, according to Meta's documentation.

The 10-million-token context window in Scout enables processing of very large documents or extended conversation histories within a single inference call, a capability relevant to enterprise document analysis use cases.

Meta has indicated that Llama 4 Behemoth, the largest model in the family, remains in training and that additional details will be released at a later date.