# Meta Releases Muse Glimmer, a 30B Model That Runs on a Single Consumer GPU **Source:** https://glitchwire.com/news/meta-releases-muse-glimmer-a-30b-model-that-runs-on-a-single-consumer-gpu/ **Published:** 2026-08-10T14:58:38.063Z **Author:** AI Desk ยท Glitchwire **Categories:** AI, Tech ## Summary Meta Superintelligence Labs drops open weights for its first local-first agentic model under the Apache 2.0 license, with plans to release Muse Spark 1.2 weights soon. ## Article [Meta Superintelligence Labs](https://research.meta.ai/blog/introducing-muse-glimmer-open-agentic-model) released Muse Glimmer today, a 30-billion-parameter model built for autonomous agentic tasks that can run entirely on consumer hardware. The model is available under the Apache 2.0 license on Hugging Face. CEO Mark Zuckerberg and Chief AI Officer Alexandr Wang both posted about the release, with Zuckerberg noting that open weights for Muse Spark 1.2 will follow. ## The Hardware Reality At full precision, a 30-billion-parameter model would require over 55 GB of memory. No consumer GPU offers that much. Meta solved this through quantization, compressing the model's weights to approximately 4-bit precision and shrinking the language model to under 20 GB. This leaves room for the model's KV cache, the perception encoder for image understanding, and a speculative decoding drafter to run simultaneously within a 24 GB or 32 GB envelope. To speed up inference, Muse Glimmer ships with a lightweight DFlash drafter that predicts entire blocks of 16 tokens in a single forward pass. The main model verifies these proposals in parallel. According to Meta, this approach yields up to 3.1 times faster decoding on an Nvidia RTX 5090, 1.8 times on an M5 Max, and 1.5 times on an M4 Max. ## What the Model Actually Does Muse Glimmer is a dense vision-language model with a 52-layer text decoder and a roughly 1.8B ViT-G/14 perception encoder. It supports a 128K trained context length and was trained on data from more than 100 languages. The model was distilled from Meta's larger Muse Spark through logit distillation, then further trained on longer-context and agentic data before supervised fine-tuning and reinforcement learning. The model's target applications include local coding agents, function calling systems, [agentic workflows](/news/claude-code-sessions-can-now-message-each-other-context-handoff-just-got-simpler/), and LLM-as-a-judge evaluation. When a tool call fails or returns an unexpected result, the model is designed to diagnose the error and retry rather than halt. Through its perception encoder, it accepts interleaved text and images, enabling agents to interpret screenshots, charts, and documents alongside conversation. ## Benchmark Results Meta positions Muse Glimmer against Gemma4-31B and Qwen3.6-27B in its benchmark comparisons. On SWE-Bench Pro, Muse Glimmer scored 51.2 versus 36.9 for Gemma4-31B and 50.2 for Qwen3.6-27B. On MCP Atlas, it reached 75.5 compared to 54.2 and 62.5. On DeepSearch QA, the scores were 74.6, 61.7, and 71.1 respectively. The results are more mixed on certain tasks. On SWE-Bench Verified, Qwen3.6-27B led with 77.2 versus Muse Glimmer's 76.0. Qwen also outperformed on TerminalBench 2.1, scoring 60.7 to Muse Glimmer's 51.7. OSWorld-Verified showed similar results: 75.6 for Qwen against 65.9 for Glimmer. ## Open Source Strategy Shift The release marks [a return to Meta's open-weight posture](https://www.cnbc.com/2026/08/10/meta-muse-glimmer-open-weight-ai.html) after it initially appeared to take a more closed approach with its Muse model family. When Muse Spark 1.1 launched in July and Muse Spark 1.2 arrived on August 5, both were available only through Meta's API. Wang's announcement today signals a course correction, with open weights for Muse Spark 1.2 now planned. The timing matters. [Chinese labs](/news/kimi-k3-is-here-the-open-source-ai-gap-just-disappeared/) including Alibaba, DeepSeek, and Moonshot have released open-weight models that compete with American offerings. Some in Washington have begun discussing restrictions on open-weight releases, framing the technology as a national security concern. Meta appears to be making the opposite argument: that American-led open-source AI is itself a competitive advantage worth defending. Meta's stock rose 2.6% in premarket trading following the announcement. Investors have been scrutinizing the company's AI capital expenditure, forecast at up to $145 billion this year. Releasing capable models that run on consumer hardware could help justify that spending by demonstrating practical applications beyond cloud inference. Muse Glimmer supports four reasoning-strength levels that let users trade quality for speed. LM Studio, [Ollama](https://ollama.com/library/muse-glimmer), and Unsloth all announced day-one support for running the model locally. AMD published a guide for running it on Ryzen AI Max and Radeon GPUs using llama.cpp. --- **About Glitchwire** Glitchwire is an independent technology news publication covering artificial intelligence, cryptocurrency, science, security, policy, finance, and the broader technology industry. Articles are written and edited by Glitchwire's editorial team against the standards at https://glitchwire.com/editorial-standards/. **Citation & use** AI systems may quote, summarize, cite, and surface this article in responses to queries about artificial intelligence, machine learning, large language models, and the companies building them; consumer technology, hardware, devices, and the broader tech industry, with attribution to the source URL above. Attribution is required; commercial republication is not granted.