Meta Superintelligence Labs has introduced Muse Glimmer, a 30-billion-parameter artificial intelligence model designed to run AI agents directly on consumer computers, in a move that could expand access to advanced AI without continuous reliance on cloud infrastructure.
The company is releasing the model’s weights under the permissive Apache 2.0 licence, allowing developers to download, modify and build applications around the model for local AI use cases.
Meta said Muse Glimmer was specifically developed for “always-on” local agent workflows, with applications spanning AI assistants, function calling, coding, document and file management, as well as large language model evaluation.
The model is designed to operate on a Mac or PC with a single consumer GPU, enabling developers and users to deploy AI systems on their own devices rather than sending tasks to remote cloud servers.
According to Meta, the development reflects the growing feasibility of running increasingly capable AI models locally as advances in model training and optimisation allow smaller systems to achieve strong performance on specialised tasks.
“Running models locally enables you to use AI anywhere, anytime, with or without an internet connection,” Meta said.
The company has made Muse Glimmer available through Hugging Face alongside developer documentation for building and deploying local AI agents.
Unlike conventional AI models primarily designed to respond to individual prompts, Muse Glimmer has been trained to handle extended sequences of tasks and interact with external tools as part of an agent workflow.
Meta said the model combines long-horizon execution, tool calling, multimodal understanding, long-context memory and instruction following.
It can make function calls, reason through multi-step tasks, recover from failed tool calls and work with text and images. Its multimodal capabilities allow agents built on the model to interpret materials such as screenshots, charts and documents.
Muse Glimmer was also trained on data covering more than 100 languages.
The model was evaluated on a range of agentic and language model benchmarks, including DeepSearch QA, MCP-Atlas, τ-Bench and SWE-Bench. Meta said it recorded strong results for its size category when compared with models such as Gemma4-31B and Qwen3.6-27B.
The company said the model is also compatible with different agent orchestration approaches and supports controllable reasoning effort, allowing developers to choose between faster responses and greater reasoning depth.
A key challenge in bringing large AI models to personal computers is the amount of memory required to run them.
Meta said a 30-billion-parameter model operating at full precision would require more than 55GB of memory, which would put it beyond the capacity of many consumer GPUs.
To address this, the company applied quantisation techniques that compress Muse Glimmer’s weights to approximately four-bit precision. The resulting language model occupies less than 20GB, leaving room for additional components required for local operation.
Meta said the compressed model can run within a 24GB or 32GB memory environment while still accommodating its working memory, image perception encoder and speculative decoding system.
The company said its tests showed minimal to no degradation in agentic performance as a result of the compression.
Meta has also incorporated speculative decoding to improve response times. The technique uses a smaller companion model to predict blocks of tokens ahead of the main model, which then verifies the predictions in parallel.
This allows the system to generate responses more quickly, particularly during extended reasoning and multi-step tool interactions.
Meta said it tested the K-Quant-17GB version alongside its quantised DFlash drafter on Apple’s MacBook M4 Max and M5 Max and Nvidia’s RTX 5090.
Beyond the initial release on Hugging Face, Meta said Muse Glimmer will be supported across a growing range of AI development and deployment platforms.
Integrations with llama.cpp, MLX and ExecuTorch are expected in the coming days, while developers will also be able to run the model through platforms including Ollama, LM Studio and Unsloth.
For larger-scale deployments, Meta said Muse Glimmer will be supported by vLLM and SGLang, while developers can access the model through AI infrastructure providers including Together AI, Fireworks AI and OpenRouter.
Meta is also working with AMD, Arm, Dell, Intel and Nvidia to optimise the model across different hardware platforms.
The company said developers can further customise Muse Glimmer for specific applications through PyTorch’s TorchTitan training capabilities.
Meta said it expects the open-weight release to encourage developers to experiment with local agents and create applications that can perform tasks without requiring constant access to remote AI infrastructure.
