Muse Glimmer is the new model of artificial intelligencefromMeta designed to function locally. It has 30 billion parameters and is designed to run AI agents directly on a computer.
The model is already available with open weights under an Apache 2.0 license. Meta aims for developers to use it without relying on a permanent internet connection.
Muse Glimmer is the new artificial intelligence model from Meta
What is Muse Glimmer and what is it for
Muse Glimmer is oriented towards tasks that require agents capable of executing actions autonomously. Its uses include programming, calling tools, and evaluating other models.
The model can also complete tasks from start to finish, work with code, and handle requests that require multiple steps. Additionally, it was trained to recover from errors when a tool fails or returns an unexpected result.
The main difference of Muse Glimmer: it works locally
Meta designed Muse Glimmer to run on a Mac or PC with a single consumer GPU. This allows the use of AI agents even without an internet connection.
The main difference of Muse Glimmer: it works locally
The model has 30 billion parameters but uses quantization techniques to reduce its size. With a precision close to 4 bits, it occupies less than 20 GB of memory.
Meta indicated that this configuration allows it to run on machines with 24 GB or 32 GB of memory, leaving space for other necessary components during its operation.
What is Muse Glimmer and what is it used for
What the new artificial intelligence model from Meta can do
Muse Glimmer was developed to work with agents that need to combine different capabilities during the same task.
Complete tasks from start to finish: works on agent benchmarks like DeepSearch QA, MCP-Atlas, τ-Bench, and SWE-Bench.
Use tools: executes function calls through defined schemas.
Reason over multiple steps: maintains plans in long and complex processes.
Recover from errors: can detect failures and retry an action.
Process images: interprets text and images, including documents, graphics, and screenshots.
Control the level of reasoning: allows choosing different levels of effort depending on the task.
Work in multiple languages: was trained with data from over 100 languages.
How Meta made Muse Glimmer work on a computer
Meta combined a compact architecture with a distillation technique that transfers reasoning capabilities from a larger teacher model.
The training was divided into three stages: pre-training, intermediate training, and post-training. In this last phase, supervised fine-tuning, distillation, and reinforcement learning were combined.
Additionally, the company applied specific optimizations to reduce memory consumption and improve response speed on consumer hardware.
The technology that accelerates Muse Glimmer
The model uses speculative decoding to generate text faster. An auxiliary model proposes blocks of tokens, and the main model verifies those proposals in parallel.
The technology that accelerates Muse Glimmer
Meta claims that this technique allows for faster generation during conversations and tasks with multiple steps, without altering the quality of the response.
The company measured the performance of a quantized version of 17 GB along with an auxiliary DFlash model on machines with MacBook M4 Max, M5 Max, and RTX 5090.
How to download and use Muse Glimmer
Muse Glimmer is already available, and its weights can be downloaded from Hugging Face. Meta is also preparing integrations with tools used to run AI models locally.
How to download and use Muse Glimmer
These include Ollama, LM Studio, Unsloth, llama.cpp, ExecuTorch, and MLX. For larger-scale implementations, vLLM and SGLang are also mentioned.
The model can also be used through services like Together AI, Fireworks AI, and OpenRouter. Meta is also working with AMD, Arm, Dell, Intel, and NVIDIA to optimize its performance on different devices.
