Technology Desk: Meta has introduced a new AI model Muse Glimmer from its Meta Superintelligence Labs. According to the company, this is an AI model with 30 billion parameters, which has been specially designed for such local AI agents, which can work continuously i.e. always on.
Meta says Muse Glimmer can be run on Macs or PCs with a single consumer GPU. It can be used for local coding, getting work done through tools, multi-step reasoning, and other agentic AI tasks.
The company has made the VATS of this model available under Apache 2.0 license. Users will be able to download its model weights and run it on their system.
Muse Glimmer has 30 billion parameters. If understood in simple language, parameters reflect the ability of the AI model through which it understands the input and generates answers.
Meta has optimized this model specifically for local agent workflows. This means that AI models will not need to depend on cloud servers or internet connections for every small task.
According to the company, Muse Glimmer can be used for tasks like agentic task completion, function calling, local coding and LLM-as-a-judge evaluation.
A major advantage of using Muse Glimmer as a local AI model is that it can be run with or without an internet connection.
Meta says many existing AI applications rely on cloud infrastructure and network connections. In such a situation, their use may be affected if internet or cloud server is not available.
Muse Glimmer is designed to mitigate such limitations. By running it locally on the computer, some AI tasks can be performed directly on the device.
The original model of Muse Glimmer is quite large. According to Meta, it may require more than 55GB of memory to run at full precision.
However, the company has used 4-bit quantization to reduce the size of the model. Through this the size of the model has been reduced to less than 20GB.
Because of this, this model may be easier to run on systems with 24GB or 32GB memory. That means users who have a consumer GPU with enough VRAM can use it locally.
Meta also compared the Muse Glimmer to models like Google's Gemma4-31b and Alibaba's Qwen3.6-27B.
According to the test results shared by the company, Muse Glimmer performed better than these two models in 12 tests.
However, these results are based on tests conducted by Meta and the benchmarks shared by it. Model performance may vary depending on different hardware and actual use.
Meta hasn't designed Muse Glimmer to just answer general questions. The company wants to make it more useful for agentic AI tasks.
Through this, AI can try to complete multi-step tasks. The model is specifically designed to:
End-to-End Agentic Task Completion
Using tools reliably
function calling
Multi-Step Reasoning
multimodal reasoning
error diagnosis
tool retry
local coding
LLM-as-a-judge evaluation
Such AI agents may try to complete a task in multiple stages based on user instructions.
According to Meta, Muse Glimmer has been trained on data from more than 100 languages. This is expected to help the model understand and work with inputs in different languages.
The company has also prepared it for multimodal reasoning. That means the model will not be limited to just text and has been designed to work with different types of inputs.
Meta has provided Muse Glimmer model vests at Hugging Face. With this, developers and AI researchers will be able to download the model and test it on their local system.
The company has said that in the coming days, Muse Glimmer will also be made available on platforms and apps like Ollama, LM Studio and Unsloth.
This can make it easier to run models locally for users who prefer to use off-the-shelf tools rather than the command line or complex AI setup.
Meta said it is working with several technology companies to improve the performance of Muse Glimmer on different devices and hardware.
The company has collaborated with AMD, Arm, Dell, Intel and Nvidia to optimize the model on different hardware.
Apart from this, Meta has also informed about working on optimized integration for llama.cpp, MLX and ExecuTorch. According to the company, these integrations will be made available in the coming days.
Model: Meta Muse Glimmer
Developer: Meta Superintelligence Labs
Parameters: 30 billion
Target use: Local and always-on AI agents
Full precision memory: more than 55GB
Size after 4-bit quantization: less than 20GB
System: Mac or PC with single consumer GPU
Memory: Designed to run on 24GB/32GB setup
License: Apache 2.0
Languages: Training on data from more than 100 languages
Model Weights: Available at Hugging Face
Upcoming Platforms: Ollama, LM Studio and Unsloth
Optimization Partners: AMD, Arm, Dell, Intel and Nvidia
Other integrations: llama.cpp, MLX and ExecuTorch
