Meta Drops Muse Glimmer, a 30B Open Agent That Runs Offline
Meta open-sources Muse Glimmer, a 30B agentic model distilled from Muse Spark 1.2 that runs on a single consumer GPU under Apache 2.0
- Free & open: Meta releases Muse Glimmer 30B weights under Apache 2.0, available now on Hugging Face.
- Runs locally: Fits on 18 GB VRAM after quantization; tested on M4 Max, M5 Max, and RTX 5090 -- no cloud required.
- Distilled from Muse Spark 1.2: Uses logit distillation and RL post-training to transfer agentic reasoning from a much larger teacher model.
- DFlash speculative decoding: A block-diffusion drafter model delivers up to 3.1x faster generation on RTX 5090 with no quality loss.
- Agent-first design: Trained for multi-step tool use, failure recovery, multimodal input, and long-horizon task completion -- not retrofitted from a chat model.
- More open weights coming: Meta has signaled open weights for Muse Spark 1.2 are also on the way.
Meta Superintelligence Labs just dropped Muse Glimmer, a 30-billion-parameter open-weight model purpose-built for agentic workflows that run entirely on your own hardware. No cloud, no API key, no internet required. The weights are available now on Hugging Face under a permissive Apache 2.0 license, free to use, modify, and build on.
The problem it solves
Foundation models have achieved remarkable capabilities across reasoning, code generation, and tool use, yet most deployments still depend on cloud infrastructure and network access. Running models locally enables you to use AI anywhere, anytime, with or without an internet connection. The catch has always been that capable models are enormous. Muse Glimmer is Meta's answer: a model that is genuinely capable at agentic tasks while fitting on a single consumer GPU.
Muse Glimmer is a distilled version of Meta's Muse Spark 1.2 model, designed with a focus on efficiency to minimize system requirements. It is the first open model from Meta Superintelligence Labs, and is released under the Apache 2.0 license. Meta's Chief AI Officer Alexandr Wang framed it simply: "a 30B model that runs on a single consumer GPU, with open weights for a version of Muse Spark 1.2 coming soon -- two very different models, both headed into people's hands."
What it can actually do
Muse Glimmer is optimized for always-on local agent workflows and is small enough to run on a Mac or PC with a single consumer GPU, enabling use cases that range from local agents and function calling, to local coding, and LLM-as-a-judge evaluation. The demo Meta published shows the model autonomously discovering a local Home Assistant instance via network tool calls, querying device APIs, writing a full HTML/CSS/JS dashboard from scratch, and deploying a local server to verify it -- all from a single natural language prompt.