Meta Releases Muse Glimmer, a 30 Billion Parameter AI Model That Runs on a Single Consumer GPU
Meta Superintelligence Labs released Muse Glimmer on August 10, a 30-billion-parameter open-weight AI model designed to run locally on consumer hardware including PCs and Macs with a single GPU. The model is available on Hugging Face under the Apache 2.0 license and supports over 100 languages.
Meta Superintelligence Labs released Muse Glimmer on August 10, 2026, a 30-billion-parameter open-weight AI model built to run on consumer hardware without requiring cloud infrastructure.
The model is optimized for local execution on PCs and Macs equipped with a single consumer GPU. Through 4-bit quantization, the model size is reduced to under 20 GB, allowing it to fit within a 24 GB or 32 GB memory envelope.
Meta integrated a speculative decoding companion network called DFlash to improve response speeds. DFlash predicts blocks of 16 tokens in parallel, which the main model then verifies. On an Nvidia RTX 5090, this produces a 3.1x speedup. On an Apple M5 Max, the speedup is 1.8x.
Muse Glimmer is designed for agentic tasks, meaning it can take sequences of actions to complete goals rather than just answering single questions. It can invoke tools with precise schemas, diagnose and retry failed tool calls, and process interleaved text and images including charts, screenshots, and documents.
The model supports over 100 languages and is compatible with frameworks including Ollama, LM Studio, llama.cpp, MLX, vLLM, and SGLang. It is available on Hugging Face under the Apache 2.0 license.
Meta is collaborating with hardware partners including AMD, Arm, Dell, Intel, and Nvidia to optimize performance across a wider range of consumer devices.
The release positions Muse Glimmer against competitors like Gemma4-31B and Qwen3.6-27B on benchmarks including SWE-Bench and MCP-Atlas.
