{"data":{"id":2166,"lang":"en","title":"Meta Open-Sources Muse Glimmer, a 30B AI Model That Runs on One Consumer GPU","summary":"Meta released Muse Glimmer, a 30B-parameter open-weight AI model that runs on a single consumer GPU under Apache 2.0.","category":"ai","category_label":"AI","published_at":"2026-08-11T02:01:19Z","updated_at":"2026-08-11T02:01:19Z","url":"https://techtools.cz/tools/news/article/2166-meta-open-sources-muse-glimmer-a-30b-ai-model-that-runs-on-one-consumer-gpu","api_url":"https://techtools.cz/news-api/v1/articles/2166?lang=en","markdown_url":"https://techtools.cz/news-api/v1/articles/2166.md?lang=en","image_url":"https://www.phoronix.net/image.php?id=2026\u0026image=muse_glimmer","word_count":150,"reading_time_minutes":1,"sources":[{"url":"https://research.meta.ai/blog/introducing-muse-glimmer-open-agentic-model","domain":"research.meta.ai"},{"url":"https://www.marktechpost.com/2026/08/10/meta-ai-releases-muse-glimmer/","domain":"marktechpost.com"},{"url":"https://www.phoronix.com/news/Meta-Muse-Glimmer","domain":"phoronix.com"}],"translation":{"lang":"cs","id":2239,"title":"Meta zveřejnila jako open source model Muse Glimmer, 30 miliard parametrů běžících na jedné běžné grafické kartě","url":"https://techtools.cz/tools/news/clanek/2239-meta-zverejnila-jako-open-source-model-muse-glimmer-30-miliard-parametru","api_url":"https://techtools.cz/news-api/v1/articles/2239?lang=cs"},"content":"Meta Superintelligence Labs released Muse Glimmer on August 10, a 30-billion-parameter open-weight AI model built to run locally on a single consumer GPU instead of cloud servers. The model is available under the permissive Apache 2.0 license on Hugging Face.\n\nMeta compressed the model's memory footprint from 55GB down to under 20GB using 4-bit quantization, letting it fit inside a 24-32GB VRAM envelope alongside its KV cache and perception encoder. A speculative-decoding technique called DFlash pushes generation speed up to 3.1x faster on an RTX 5090 and 1.8x faster on a Mac M5-Max compared to standard token-by-token decoding.\n\nDistilled from Meta's larger Muse Spark model, Glimmer targets agentic workflows: multi-step reasoning, tool use, function calling, coding, and document/screenshot interpretation across more than 100 languages. Meta says it beats similarly sized rivals Gemma4-31B and Qwen3.6-27B on agentic, coding, reasoning, and multimodal benchmarks, sharpening the open-weight competition with OpenAI and Anthropic."}}