Meta presents Muse Glimmer, its new open model with 30 billion parameters that runs on the device

Meta has presented its new open model Muse Glimmer that, with 30 billion parameters, it is executed in the form local on both computers and a single consumer GPU, without the need for a connection to the cloud.

This is the first Meta Superintelligence Labs model that has been published with open weights, low Apache 2.0 license, and is designed to solve complex multi-step tasks autonomously.

Specifically, Muse Glimmer is optimized for always-on local agent workflows, as the company explained in a statement. Likewise, by having 30,000 million parameters, it is “small enough” to run on a PC with a single consumer GPU.

This means that it is capable of executing use cases that span from local agents and role calls to local coding and LLM assessment ‘as a judge’. In practice, this includes agents managing a calendar, composing messages, organizing files, and learning from the user’s way of working.

That is, it can perform long-term execution tasks with multi-step reasoning, precise tool selection, multimodal understanding, failover, context memory extensive and following instructions.

However, it stands out notably for being able to do so locally anywhere, anytime, with or without an internet connection and with high speed so as not to interrupt the workflow. Thus, they have used quantization techniques to compress the model weights to an approximate precision of 4 bits, reducing its size to less than 20 GB.

Its size and performance are possible because, as he explained, smaller models, when trained effectively, can “achieve cutting-edge performance on specific tasks,” as Meta has assured.

Specifically in this case, Meta has trained Muse Glimmer to balancing capacity with local hardware processing and memory limitations, with a “novel distillation recipe” that is based on transfer agent reasoning from a “much larger” master model” and inference optimizations to “meet latency expectations.”

Thus, the company has specified that it has trained with the Muse Spark results, more extensive context data, and a distillation mix based on policies and reinforcement learning in general domains of reasoning, coding and agents. It also has multilingual capabilities by training with data from more than 100 languages.

In the evaluation tests provided by Meta, the company has compared to other models such as Google’s Gemma4-31B and Alibaba’s Qwen3.6-27B, where he has obtained a solid performancefor its size, surpassing these models in ‘benchmarks’ of general agentic tasks, agentic coding and general and reasoning abilities, among others.

All in all, Muse Glimmer is now available and can be downloaded at Hugging Face. Likewise, in the coming days, it will be possible to run locally with Meta partners such as Ollama, LM Studio and Unsloth. It can also be deployed with frameworks such as llama.cpp, ExecuTorch and MLX, among others.

In addition to all this, Meta has announced that soon will open the weights of a version of Muse Spark 1.2, one of its most advanced foundational models.