{"rewrite":{"id":"r_c2aaa0c02b5328399ae72bd0","clusterId":"c_6d8c6e91f4bce928dba67157","slug":"meta-releases-open-model-muse-glimmer-with-29-6b-parameters","model":"deepseek-v4-flash:free","headline":"Meta Releases Open Model Muse Glimmer With 29.6B Parameters","summary":"Meta released Muse Glimmer, an open AI model with 29.6 billion parameters, on August 10, 2026. The dense model handles text and image inputs. Benchmarks show it outperforming Google's Gemma 4 31B across many tests. It supports the DFlash speculative decoding method for faster local processing.","whyItMatters":"Muse Glimmer shows a local, open model at 29.6B parameters outperforming comparable closed and open rivals on benchmarks, narrowing the gap between device-runnable AI and cloud models.","webCardHtml":"\u003cp\u003eMeta positioned Muse Glimmer as a device-runnable model with competitive performance. The company published benchmark tables comparing it against Gemma 4 31B and Qwen3.6-27B, with Muse Glimmer leading in most tests.\u003c/p\u003e\u003cp\u003eThird-party evaluator Artificial Analysis also tested the model. Its composite index places Muse Glimmer above the closed Claude Haiku 4.5 and close to Gemini 3.5 Flash-Lite. Meta described the result as a large jump over the Llama 4 series from 2025.\u003c/p\u003e\u003cp\u003eThe model uses DFlash, a speculative decoding method that generates drafts with a lightweight diffusion model. On an NVIDIA GeForce RTX 5090, DFlash raises decode speed by 3.1 times to 233 tokens per second. Weights fit in 24GB of VRAM, and the license is Apache License 2.0. LM Studio and Ollama already support it.\u003c/p\u003e","blueskyPost":"Meta released Muse Glimmer, a 29.6B open model that runs locally. Benchmarks show it ahead of Gemma 4 31B in most tests and above Claude Haiku 4.5 on the Artificial Analysis index. Apache 2.0, fits in 24GB VRAM.","twitterPost":"Meta's Muse Glimmer: a 29.6B open model for local use. It beats Gemma 4 31B on most benchmarks, scores above Claude Haiku 4.5 on Artificial Analysis, and hits 233 tokens/sec on an RTX 5090 with DFlash. Apache 2.0.","threadsPost":null,"newsletterBlurb":"Meta released Muse Glimmer, a 29.6B-parameter open model built for local execution. Benchmarks show it ahead of Google's Gemma 4 31B in most tests, and the Artificial Analysis index places it above Claude Haiku 4.5. DFlash decoding reaches 233 tokens per second on an RTX 5090, and the Apache 2.0 weights fit in 24GB of VRAM.","attributionJson":"[{\"source\":\"GIGAZINE\",\"url\":\"https://gigazine.net/news/20260811-meta-muse-glimmer/\",\"title\":\"Metaがローカル動作するオープンモデル「Muse Glimmer」を公開、30BでGoogleのGemma 4 31Bより高性能\"}]","lintFlagsJson":null,"lintHits":0,"costUsd":0,"inputTokens":5199,"outputTokens":690,"status":"published","repairAttempts":0,"nextRepairAt":null,"factsAttemptedAt":1786434202,"createdAt":"2026-08-11T07:33:30.000Z","publishedAt":"2026-08-11T07:36:44.000Z","updatedAt":"2026-08-11T07:33:30.000Z"},"cluster":{"id":"c_6d8c6e91f4bce928dba67157","canonicalTitle":"Metaがローカル動作するオープンモデル「Muse Glimmer」を公開、30BでGoogleのGemma 4 31Bより高性能","representativeArticleId":"a_e63bf85ce4dd4e2e80fe6855","sourceCount":1,"writtenSourceCount":1,"writeAttempts":0,"isSolo":true,"entitiesJson":"{\"anime_titles\":[],\"manga_titles\":[],\"work_titles\":[\"Muse Glimmer\"],\"studios\":[],\"people\":[],\"type\":\"news\",\"domain\":\"other\",\"is_roundup\":false}","contentType":"news","status":"published","firstSeenAt":"2026-08-11T01:58:00.000Z","lastSeenAt":"2026-08-11T01:58:00.000Z","updatedAt":"2026-08-11T07:36:44.000Z"},"attribution":[{"source":"GIGAZINE","url":"https://gigazine.net/news/20260811-meta-muse-glimmer/","title":"Metaがローカル動作するオープンモデル「Muse Glimmer」を公開、30BでGoogleのGemma 4 31Bより高性能"}],"entities":{"anime_titles":[],"manga_titles":[],"work_titles":["Muse Glimmer"],"studios":[],"people":[],"type":"news","domain":"other","is_roundup":false},"keyFacts":null}