Meta releases Muse Glimmer AI model designed to run on a single computer

Meta has introduced Muse Glimmer, a smaller artificial intelligence model designed to run on a single computer rather than relying on large cloud-based systems. The company said the new model is intended for agent-oriented tasks, including scheduling, managing files and other activities that require an AI system to interact with tools and complete several steps.

Muse Glimmer is based on Meta’s Spark 1.2 model, though the new version is significantly smaller. This allows it to operate using a single graphics processing unit (GPU), making it more accessible to developers and users who want to run AI models on local hardware. Meta said the model was designed with the limitations of consumer and local computing systems in mind.

“We designed Muse Glimmer to balance capability against the memory and compute constraints of local hardware,” Meta wrote when announcing the release. The company said the approach is intended to make capable AI agents easier to operate without depending entirely on remote servers or large-scale data centre infrastructure.

The release comes as interest grows in smaller AI models that can operate locally. Running an AI model on a personal computer can give users greater control over their data and reduce the need to send information to external cloud services. It can also make AI applications more practical in situations where an internet connection is unreliable or where low-latency responses are important.

Model weights made available to developers

Meta said it is making the weights used by Muse Glimmer available to the public on Hugging Face, along with documentation to help developers work with the model. The download is available free of charge, allowing users to obtain the model and run it on compatible personal computers.

The company also said optimised integrations are expected to become available through llama.cpp and other platforms. These integrations are intended to simplify the process of setting up the model and connecting it to applications that can use AI agents. Meta said users should be able to “go from download to working agent in minutes” once those integrations are available.

Despite its smaller size, Meta claims that Muse Glimmer can deliver strong results across several AI benchmarks. The company highlighted performance on DeepSearch QA, MCP-Atlas and SWE-Bench, the latter of which is designed to assess how effectively an AI model can write, understand and debug computer code.

Muse Glimmer also supports several capabilities aimed specifically at agent-based applications. These include tool use, multi-step reasoning, and recovery from task failures. The model supports multimodal input and is designed to work with various scaffolding systems for building AI agents. Meta also said the model can work with OpenClaw and other agent orchestrators.

The company added that Muse Glimmer was trained using data covering more than 100 languages. That could make the model useful to developers building applications for users across different regions. However, its real-world performance will depend on the specific hardware, software configuration and applications in which it is deployed.

Meta embraces a more open AI strategy

Muse Glimmer situates Meta within a broader shift towards smaller AI systems that users can download and run on their own hardware. Several Chinese AI companies, including DeepSeek, have gained attention for models that can run locally and are offered under comparatively open licensing arrangements.

Meta’s release also reflects a broader debate over how advanced AI should be distributed. Rather than concentrating increasingly capable models within a small number of cloud providers, proponents of local AI argue that developers and individuals should have greater access to models that can be operated independently.

“Rather than centralising superintelligence, we should distribute it widely and give every person the ability to direct it,” Meta chief executive Mark Zuckerberg said in an essay accompanying the Muse Glimmer release.

The move could also help Meta strengthen its position in an increasingly competitive AI market. Its Muse Spark AI system has faced criticism for its capabilities compared with those of models from leading competitors such as OpenAI and Anthropic. By releasing a smaller model that can operate locally, Meta appears to be placing greater emphasis on accessibility and developer adoption rather than competing solely through the size of its cloud-based systems.

The strategy comes as local and open AI models attract increasing attention from the technology industry. Hardware companies and AI developers are exploring ways to make capable models more efficient, while efforts are also underway to address security risks associated with openly available AI systems.

Nvidia recently launched the Open Secure AI Alliance, an initiative focused on improving cyber defence and security around AI systems. The development highlights the growing tension between making AI models widely available and ensuring that the software and infrastructure surrounding them remain secure.

Muse Glimmer therefore represents more than another model release for Meta. Its ability to run on a single GPU, combined with publicly available model weights and developer documentation, could make it an option for users who want to experiment with AI agents without relying entirely on cloud-based services.

Meta is betting that this combination of relatively small hardware requirements, agent-focused capabilities, and broader access can attract developers seeking more control over where and how their AI systems operate.