Meta Superintelligence Labs on Monday released Muse Glimmer, a 30-billion-parameter open-weight language model designed to run entirely on a single consumer graphics card — Meta's first major open model release in more than a year.
A model of that size would typically need roughly 55 gigabytes of memory to run. Meta's engineers compressed it to under 20 gigabytes using four-bit quantization and a "drafter" system that generates a fast initial response before a verification step refines it, according to a review of the release by SiliconANGLE. The result, Meta says, is a model that fits on a single gaming PC or Mac while still handling multi-step agentic tasks such as coding, research and tool use.
Muse Glimmer was distilled from Meta's larger Muse Spark model, then put through additional training passes aimed at reasoning and agentic performance, and it ships with a 2-billion-parameter vision component feeding a 28-billion-parameter text decoder. Meta says the model outperforms similarly sized open rivals, including Gemma4-31B and Qwen3.6-27B, on roughly half of two dozen benchmarks it was tested against, with particular strength in coding, web research and reading scientific charts, per Phoronix's coverage. The weights are released under an Apache 2.0 license and are already available through Hugging Face, with day-one support in Ollama, LM Studio, llama.cpp, MLX, vLLM and SGLang.
A policy essay attached to the release
Meta paired the technical release with a lengthy essay from chief executive Mark Zuckerberg arguing against concentrating advanced AI capability inside a small number of companies or governments, and framing open-weight releases like Muse Glimmer as central to that strategy.
Rather than centralizing superintelligence, we should distribute it widely and give every person the ability to direct it.
Mark Zuckerberg, Meta
The essay also urges Washington to avoid regulation that would make it harder to build open models domestically, warning that Chinese developers — which have released a steady stream of competitive open-weight models over the past year — could otherwise take the lead in the category. It's a notable pivot for Meta, which had pulled back on major open releases after facing criticism over the safety profile of earlier Llama models and as competition intensified from Chinese labs such as Alibaba and DeepSeek.
Meta says it now evaluates every model against internal safety criteria before deciding how broadly to release it, a governance step the company added since its last open release. Whether Muse Glimmer's low-memory, single-GPU profile becomes the new competitive battleground for open models — rather than raw benchmark scores — may become clearer as rivals respond in the coming weeks.
