Meta has launched Muse Glimmer, a 30-billion-parameter model designed to run on consumer GPUs for local AI agent tasks.
Key facts
- •Muse Glimmer is a 30-billion-parameter model released under an Apache 2.0 license.
- •The model uses 4-bit quantization to run on consumer GPUs with under 20 GB of memory.
- •Meta reports the model scored 75.5 on the MCP Atlas benchmark, outperforming Gemma4-31B and Qwen3.6-27B.
- •A DFlash-based drafter is used to speed up token generation by proposing blocks of tokens for parallel verification.
- •The model supports OpenClaw and other orchestration patterns for managing local agent tasks.
Meta has released Muse Glimmer, a 30-billion-parameter AI model, under an Apache 2.0 license. The model is designed to operate on consumer GPUs, allowing developers to run local AI agents that manage private data such as schedules, files, and messages without relying on cloud-hosted infrastructure.
By the numbers
Performance and Benchmarking
Meta compared Muse Glimmer against Gemma4-31B and Qwen3.6-27B across various benchmarks. Muse Glimmer outperformed competitors on several agentic tests, including MCP Atlas with a score of 75.5 and DeepSearch QA at 74.6. In coding evaluations, it led SWE-Bench Pro with a score of 51.2. However, Qwen3.6-27B demonstrated superior performance in other areas, such as OSWorld-Verified, where it scored 75.6 compared to Muse Glimmer’s 65.9.
Technical Specifications and Hardware
To fit on consumer hardware, Muse Glimmer utilizes approximately 4-bit weight quantization, reducing the model size to under 20 GB. This allows the model, alongside a perception encoder and a DFlash-based speculative-decoding drafter, to fit within a 24 GB or 32 GB memory envelope. Meta tested the model on MacBook M4-Max, MacBook M5-Max, and RTX-5090 hardware, reporting that the DFlash drafter accelerates token generation while maintaining output quality.
Safety and Integration
The model features a perception encoder capable of processing interleaved text and images, enabling agents to interpret visual inputs like screenshots and charts. Meta also conducted safety evaluations, including CI Memories and Siren AgentDojo, noting varying violation and utility rates. The model weights are available on Hugging Face, with planned integrations for llama.cpp, MLX, and ExecuTorch.
Advertisement
This article was independently rewritten by ManyPress editorial AI from reporting originally published by Artificial Intelligence News.
