What Llama is best at
- Privacy-critical work - nothing ever leaves your machine
- Builders fine-tuning a model for one specific job
- Running a real agent offline: Muse Glimmer fits on one consumer GPU
Where it isn’t the answer
People who just want the best answer with zero setup - a hosted frontier model will beat your local box.
Plans & pricing
The story so far
- Feb 2023LLaMA leaks, accidentally igniting the open-model era
- Jul 2024Llama 3.1 405B - open weights go frontier-scale
- Apr 2025Llama 4 Scout and Maverick bring mixture-of-experts to the family
- Apr 2026Meta Superintelligence Labs debuts Muse Spark, the new flagship line
- Jul 2026Muse Spark 1.1 opens up through the Meta Model API
- Aug 5, 2026Muse Spark 1.2 ships alongside Muse Code, a terminal coding agent
- Aug 2026Muse Glimmer: a 30B Apache-2.0 agent model that runs on 24 GB of VRAM
- Sep 2026Muse Spark 1.3 adds long-horizon work and asks clarifying questions before it acts
Questions people actually ask
Is there a Llama 5?
No. Meta’s flagship open-weight work moved to the Muse family under Meta Superintelligence Labs - Muse Spark for the big model, Muse Glimmer for the one you can actually run at home. Llama 4 remains the last Llama-branded release.
Is Llama really free?
The weights are free to download and use under Meta’s community licence, including commercially for most companies. Muse Glimmer goes further and ships under Apache 2.0.
How do I run it locally?
The easy path is Ollama: one install, one command, and a Llama or Muse model runs on your own machine - no account, no cloud.
Is it as good as ChatGPT?
The top open models are close on many tasks but generally a step behind the frontier. The trade is control and privacy for peak quality.