Llama 4

Meta's open-source Llama 4 family and the proprietary Muse Spark, leading the open-weights movement and closing the gap with frontier models.

LlamaMetaOpen SourceLanguage ModelLarge Language ModelAI AssistantMultimodalMoE
Developer
Meta
Type
Multimodal Language Model
License
Open Source / Proprietary (Muse)

Overview

The Llama 4 family, released on April 5, 2025, represents the final chapter in Meta's foundational open-weights era. Featuring the Scout and Maverick series, Llama 4 introduced the Mixture-of-Experts (MoE) architecture to the open-source community at a massive scale.

As of April 2026, Meta has transitioned its frontier AI efforts to a new, proprietary model series called Muse Spark. Developed by the newly formed Meta Superintelligence Labs, Muse Spark represents a strategic shift toward closed-source, state-of-the-art models designed to compete directly with GPT-5.4 and Claude 4.7 Opus. While the Llama 4 models remain the industry standard for open-source deployment, the frontier of Meta's intelligence is now hosted within the Muse ecosystem.

Llama 4 Scout (The Efficient Long-Context)

Scout is the high-efficiency multimodal model of the 2025 release. With 17B active parameters (109B total), it pioneered the 10M token context window for open models.

  • Context Window: 10 million tokens.
  • Architecture: 8 experts / 2 active.
  • Best for: Extreme RAG systems and long-video analysis.

Llama 4 Maverick (The Workhorse)

Maverick is the primary general-purpose model, featuring 17B active parameters out of 400B total (utilizing 128 experts). It was designed to rival the reasoning capabilities of the previous generation's proprietary flagships.

  • Context Window: 1 million tokens.
  • Best for: Coding, logic-heavy tasks, and professional-grade multimodal understanding.

Muse Spark (The Proprietary Frontier)

Released in early April 2026, Muse Spark is Meta's newest intelligence engine. Unlike Llama 4, Muse Spark is closed-weights and available only via Meta's secure API. It features native agentic orchestration and is optimized for the "Superintelligence" benchmarks of 2026.

Capabilities

  • State-of-the-Art Reasoning: Maverick, in particular, is designed to be a leader in complex logical and mathematical reasoning.
  • Advanced Multimodality: All models can natively process and understand both text and images.
  • Massive Context Windows: Scout's 10 million token context window opens up new possibilities for long-form content analysis.
  • Open Source: The models are freely available for both research and commercial applications, driving innovation across the AI community.

Use Cases

  • Complex Coding & Development: Maverick serves as a powerful pair programmer for developers.
  • Large-Scale Data Analysis: Scout can analyze entire codebases or vast quantities of documents in a single prompt.
  • Creative Multimodal Content: All models can be used to understand and interact with visual information.
  • Academic & Commercial Research: The open nature of the models makes them a go-to choice for researchers pushing the boundaries of AI.

Pricing & Access

  • Llama 4 (Scout/Maverick): Free to download and self-host under Meta's permissive license. Available on Hugging Face.
  • Muse Spark: Available via Meta's API for enterprise customers. Pricing is reported to be competitive with GPT-5.4 at approximately $2.50 per 1M input tokens.
  • Meta AI: Integrated into Instagram, WhatsApp, and Facebook, now powered by the Muse Spark engine.

Ecosystem & Tools

  • Official Llama Website: The primary source for downloading the models and reading the documentation.
  • Hugging Face: The main community hub for accessing and fine-tuning Llama models.
  • API Access: Available through major cloud providers and dedicated AI model hosting platforms.

Community & Resources

Frequently Asked Questions

The Llama 4 family was released on April 5, 2025, marking the last major open-weights release from Meta before their pivot to the Muse series.
Muse Spark is Meta's latest proprietary, closed-source frontier model released in April 2026 under Meta Superintelligence Labs.
No, Llama 4 Behemoth remained an internal teacher model used for distillation and was never publicly released as an open-weights model.
Llama 4 Scout features a massive 10 million token context window, designed for extreme RAG and document analysis tasks.

Explore More Models

Discover other AI models and compare their capabilities.