Overview
- Tool name: Llama 4
- Developer: Meta
- Official website: https://llama.meta.com
- Category: AI Models
- License: Llama 4 Community License (open weights, not Apache)
Llama 4 is Meta's open-weight model family released April 5, 2025 - the first Llama generation to move to natively multimodal, mixture-of-experts (MoE) architecture. The family:
- Llama 4 Scout: 17B active / 109B total parameters (16 experts), a 10-million-token context window - the largest of any public model; fits a single server GPU with 4/8-bit quantization
- Llama 4 Maverick: 17B active / 400B total (128 experts), 1M-token context, the performance workhorse
- Llama 4 Behemoth (~2T): a teacher model that was ultimately never released
Benchmarks (official): Maverick scores MMLU-Pro 80.5, GPQA Diamond 69.8, MMMU 73.4; Scout 74.3 / 57.2 / 69.4.
Where things stand in 2026
Meta has not shipped a new Llama model since April 2025; its frontier attention moved to the Muse line: Muse Spark (April 2026, closed-weights, API-only) and Muse Glimmer (August 2026, 30B, Apache 2.0, aimed at local agents on a single consumer GPU). Scout and Maverick remain freely downloadable, runnable and fine-tunable, and are still among the most common self-hosting choices - every serving engine (vLLM, llama.cpp, Ollama) and most tooling supports them natively.
License notes
Llama 4 uses the Community License, not Apache 2.0: free for commercial use, but with a 700-million-MAU threshold (above it you need a separate Meta license), naming and attribution requirements ("Built with Llama", derived model names must start with "Llama"), and an EU carve-out for multimodal models. Strictly speaking it is open weights, not OSI open source.
Access and pricing
- Weights: free download from Hugging Face / Meta; self-hosting costs compute only
- Cloud APIs: Groq, Together AI, Replicate, AWS Bedrock and others, typically $0.10-$0.49 per million input tokens
- Scout is free to try at Meta.ai
Who it is for
- Very long context: Scout's 10M window suits whole codebases, legal archives and long histories
- Private deployment: the most mature tooling ecosystem of any open-weight family
- Fine-tuning: the richest LoRA/QLoRA ecosystem
Limitations
- Unchanged since April 2025; the open-source frontier (DeepSeek, GLM, Qwen) has moved ahead on hard reasoning
- The non-Apache license is a blocker for some corporate legal teams
- Behemoth never shipped, leaving the flagship tier vacant
Frequently asked questions
What is the latest Llama version?
Llama 4 (April 5, 2025), with Scout and Maverick as the downloadable models. Meta's newer Muse line is a separate family; the Llama line effectively pauses at Llama 4.
Is Llama 4 free for commercial use?
Yes under the Community License for nearly every company; products above 700 million monthly active users need a separate Meta license. Mind the attribution and naming clauses.
Scout or Maverick?
Scout for the 10M context window and single-GPU operation; Maverick for stronger general capability with a 1M window.
*This page is not affiliated with Meta. Information verified as of October 2026; see llama.meta.com for current details.*
