By Ken Muthomi - Reporter
U.S.A, August 10 , 2026 - Meta has unveiled a new AI model designed to run directly on consumer computers, potentially allowing developers to build and use AI agents without relying on cloud infrastructure or an internet connection.
The model, Muse Glimmer, was announced on Monday by Meta Superintelligence Labs and released with its weights under the permissive Apache 2.0 licence, allowing developers to download, modify and build applications around it.
Meta said Muse Glimmer is optimised for local agentic workflows, including coding, function calling, tool use and AI evaluation, abd is small enough to run on a Mac or PC which have a single consumer GPU.
Meta said the model is currently available through Hugging Face, while integrations with platforms including Ollama, LM Studio, llama.cpp, ExecuTorch and MLX are expected in the coming days.
"Keeping with our long tradition of sharing fundamental AI research, we're releasing Muse Glimmer open weights today on Hugging Face, along with developer documentation to help you start building and running your own agents,"the company said
"Muse Glimmer is built to work with the tools developers already use. Optimized integrations on llama.cpp, MLX, and ExecuTorch will land in the coming days, so you can go from download to working agent in minutes, it added
Muse Glimmer is designed to support agentic tasks that require an AI system to carry out multiple steps to complete a user's request such as managing tool calls, sustain multi-step reasoning, recover from failed tool calls and work with text and images.
The model can also interpret screenshots, charts and documents through a dedicated perception encoder, while Meta said it has been trained on data covering more than 100 languages.
More from Kenya
Meta said it trained Muse Glimmer using a combination of distillation, supervised fine-tuning and reinforcement learning. The company used outputs from its larger Muse Spark model during pre-training to transfer some of the larger model's capabilities into the smaller system.
The company also trained the model on longer-context and agent-focused data before applying additional training across reasoning, coding and agentic tasks.
"We evaluated Muse Glimmer across a broad range of benchmarks to assess the diverse capabilities required for effective autonomous agent behavior. Compared with Gemma4-31B and Qwen3.6-27B, Muse Glimmer performs strongly for its size class on several widely used LLM benchmarks,"the company stated.
One of the main challenges with running large AI models locally is the amount of memory required, with meta noting that a 30-billion-parameter model operating at full precision would require more than 55GB of memory.
To address this, the company used quantisation techniques to reduce Muse Glimmer to approximately 4-bit precision, bringing the model's size to below 20GB, which will allow the model to run within a 24GB or 32GB memory.
The company is also working with hardware companies including AMD, Arm, Dell, Intel and NVIDIA to optimise Muse Glimmer for different devices.