Meta Releases Muse Glimmer, a 30-Billion-Parameter Open-Weight Model Built to Run on a Single Consumer GPU
Meta released Muse Glimmer, a 30B-parameter open-weight model distilled from Muse Spark for local agent workflows, alongside a Zuckerberg essay on open-source AI.
Overview
Meta released Muse Glimmer on Monday, a 30-billion-parameter open-weight model distilled from its proprietary Muse Spark system and released under the Apache 2.0 license, according to MarkTechPost. The model is optimized to run on a single consumer GPU rather than in the cloud, according to Constellation Research, and arrives alongside a lengthy essay from CEO Mark Zuckerberg making the case for distributing advanced AI broadly rather than concentrating it among a small number of labs.
What We Know
Muse Glimmer is a dense model — meaning it activates all of its parameters on every token rather than routing through a mixture of experts — according to NVIDIA’s developer blog. It is multimodal, built around a dedicated perception encoder for handling images alongside text, according to MarkTechPost. It was trained on Muse Spark’s own outputs through a distillation process, according to Constellation Research, and supports a context window of more than 120,000 tokens, according to MarkTechPost and NVIDIA.
The release is squarely aimed at local, “always-on” agent workloads rather than chat. According to Engadget, the model supports tool use, multi-step reasoning, failure recovery, and multimodal input, and is compatible with agent orchestrators like OpenClaw. Meta says the model was trained on data spanning more than 100 languages, according to Engadget.
Running a 30-billion-parameter model locally normally requires more than 55 GB of memory at full precision, but Meta compresses Muse Glimmer to roughly 4-bit precision, bringing quantized variants down toward the 17-to-24 GB range suited to a single high-end consumer GPU, according to MarkTechPost. NVIDIA’s blog specifically points to its GeForce RTX 5090, a 32GB-VRAM consumer card, as a target device for the model, according to NVIDIA. Meta also ships a speculative-decoding “drafter” model, DFlash, that MarkTechPost says lifts throughput on an RTX 5090 by roughly 3.1 times, according to MarkTechPost.
On benchmarks, Meta positions Muse Glimmer against two other openly available models in its size class, Gemma4-31B and Qwen3.6-27B. On the MCP-Atlas benchmark, Muse Glimmer scores 75.5 compared with 54.2 for Gemma4-31B and 62.5 for Qwen3.6-27B; on AIME 2026 it scores 94.7 against 89.2 and 94.1, respectively; and on SWE-Bench Pro it scores 51.2, according to results reported independently by both MarkTechPost and Hugging Face’s blog. MarkTechPost characterizes the model as winning on agentic orchestration and reasoning tasks while trailing rivals on computer-use and terminal work, according to MarkTechPost.
The release lands alongside an essay Zuckerberg published the same day, titled “The Future is for Everyone: The Path to a Positive AI Future,” according to Forbes. In it, Zuckerberg argues for spreading access to powerful AI rather than concentrating it, writing that if advanced AI systems are held by a small number of institutions, it would “naturally lead to outcomes that are less favorable for everyone else,” according to Forbes. Zuckerberg struck a similar note in comments tied to the Muse Glimmer launch, saying “Rather than centralizing superintelligence, we should distribute it widely,” according to Engadget.
Muse Glimmer’s teacher model was Meta’s closed Muse Spark 1.2, according to Engadget, and Meta confirmed plans to open the weights of that model too, according to Forbes and Constellation Research.
What We Don’t Know
Meta has not published an exact release date for the Muse Spark 1.2 weights beyond the general framing that they are coming. The precise attention-head architecture and full context-window specification are described inconsistently across outlets covering the release, so this article reports only the figures independently corroborated across multiple sources. It’s also unclear how Muse Glimmer performs against proprietary closed models from OpenAI, Anthropic, and Google, since the comparisons Meta has published so far are limited to other open-weight models in a similar size class.
Analysis
Muse Glimmer is Meta’s third distinct Muse-branded release in as many months, following the Muse Code coding agent and the Muse Image generation tools. Pairing an open-weight local model with a public essay on AI concentration signals that Meta is trying to frame its open-weight strategy as a values position — not just a product choice — at a moment when rival frontier labs have largely kept their most capable models closed.