Meta Releases Muse Glimmer: Open 30-Billion-Parameter Model

Contents

Meta has introduced Muse Glimmer, a new open AI language model specifically designed for use on ordinary consumer hardware. The Artificial Intelligence-driving Meta Superintelligence Labs is releasing the weights of the 30-billion-parameter model under the Apache 2.0 license on Hugging Face – free to use for commercial purposes as well. The release was accompanied by a very long letter from Meta CEO Mark Zuckerberg, outlining his company’s future AI strategy.

Unlike many current top models, Muse Glimmer does not rely on cloud infrastructure. It is designed to run on a Mac or PC with a single consumer GPU and, according to Meta, is suitable for local agents, function calling, coding tasks, and use as an LLM judge for evaluating other model outputs. Meta states that knowledge from the larger model Muse Spark, which is to remain closed in the future, has been incorporated into the model.

Technical Tricks for Speed on Weak Hardware

To ensure the model fits into the memory of a consumer computer at all, Meta shrinks the weights to around 4-bit precision. This reduces the size from over 55 GBytes at full precision to under 20 GBytes.

Based on its own benchmarks (including SWE-Bench, τ-Bench, MCP-Atlas, DeepSearch QA), Meta aims to show that Muse Glimmer performs well within its weight class compared to similarly sized models like Gemma4-31B and Qwen3.6-27B – as usual, the reliability of such manufacturer-specific comparisons will only be proven in independent tests.

Meta announces support from Ollama, LM Studio, Unsloth, llama.cpp, ExecuTorch, MLX, vLLM, SGLang, as well as hosting offers from Together AI, Fireworks AI, and OpenRouter. With Muse Spark 1.1, Meta had previously announced aggressive prices for its commercial model API to take market share from OpenAI and Anthropic. For adaptation to specific use cases, Meta refers to Pytorch’s TorchTitan training framework. The company is collaborating with AMD, Arm, Dell, Intel, and Nvidia on hardware optimization.

Zuckerberg: Open AI Capabilities for Everyone

Zuckerberg himself places the model within Meta’s AI strategy in his concurrently published policy paper “The Path to a Positive AI Future”. The central question of his time is who will have access to superintelligence and who will control it. His answer: the broadest possible distribution of capabilities, because the more people have access to cybersecurity superintelligence, the safer all technical systems will ultimately become. Applied to Muse Glimmer, this means: open weights are not a gift from Meta, but an expression of a theory that the concentration of AI capabilities among a few actors is more dangerous than their broad dissemination.

In his long letter, Zuckerberg also attempts to allay numerous fears: AI can be useful for the job market and the economy, he argues. Actively using it can help defend against threats in cybersecurity, bioterrorism, and other harmful AI-generated products. And the concern that the USA could lose its leading role is also addressed by the Facebook founder.

What Meta Hopes to Gain

This argumentation also serves as a counter-proposal to competing providers: Zuckerberg accuses those who rely on closed models of tending to strengthen institutions rather than individuals – most others focus on AI for businesses, governments, or other institutions. Meta, on the other hand, positions itself as the company that specifically builds superintelligence for individuals. Behind this is likely the primary intention to be the company that sets the standards in the future.

(mki)