--- title: Muse Glimmer 30B Architecture Notes | daily.dev description: Meta released Muse Glimmer 30B, a 30B multimodal reasoning model with a Gemma-like architecture. Key architectural highlights include a 131k context window,... canonical: https://daily.dev/posts/muse-glimmer-30b-architecture-notes-dkdmm5p34 twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Muse Glimmer 30B Architecture Notes | daily.dev og:description: Meta released Muse Glimmer 30B, a 30B multimodal reasoning model with a Gemma-like architecture. Key architectural highlights include a 131k context window,... og:url: https://daily.dev/posts/muse-glimmer-30b-architecture-notes-dkdmm5p34 og:image: https://api.daily.dev/og/posts/dKdmM5P34.png og:image:alt: Muse Glimmer 30B Architecture Notes og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Muse Glimmer 30B Architecture Notes **[Sebastian Raschka](https://daily.dev/sources/sebastianraschka)** · 3 min read · 0 upvotes · 0 comments ## Summary Meta released Muse Glimmer 30B, a 30B multimodal reasoning model with a Gemma-like architecture. Key architectural highlights include a 131k context window, dense (non-MoE) design, hybrid attention with a 3:1 SWA:GQA ratio, gated attention for both attention types, and an extreme GQA ratio of 32 query heads to just 2 KV heads. This results in an exceptionally small KV cache of 52 KiB per token in BF16 — far smaller than Qwen3.6 27B (64 KiB) or Gemma 4 31B (840 KiB). Benchmark performance is roughly on par with Qwen3.6, with Meta's own benchmarks showing it slightly ahead and independent benchmarks slightly behind. The model is particularly well-suited for agentic workflows due to its low memory footprint and fast prefill/decode speed. Muse Spark, the more capable model it was distilled from, remains API-only. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### What is the KV cache size per token for Meta Muse Glimmer 30B compared to Qwen3.6 27B and Gemma 4 31B? Muse Glimmer 30B has a KV cache of 52 KiB per token in BF16, compared to 64 KiB for Qwen3.6 27B and 840 KiB for Gemma 4 31B. This extreme efficiency comes from its 32 query heads to just 2 KV heads GQA ratio, making it particularly memory-efficient for long-context and agentic workloads. _Engineers choosing models for memory-constrained agentic deployments track comparisons like this on daily.dev._ ### What is the SWA to GQA ratio in Meta Muse Glimmer 30B and how does it compare to Gemma 4? Muse Glimmer 30B uses a 3:1 local-to-global (SWA:GQA) ratio in its hybrid attention design. Gemma 4 uses a 5:1 ratio for comparison. Glimmer also applies gated attention — a sigmoid gate on the attention output — to both its SWA and GQA layers, which has become increasingly common in recent model releases. _Researchers tracking LLM architectural trends across model releases find those discussions on daily.dev._ ## Similar posts on daily.dev - [Muse Glimmer 30B runs on my 32GB GPU, handles its own tool calls, and never leaves the machine](https://daily.dev/posts/muse-glimmer-30b-runs-on-my-32gb-gpu-handles-its-own-tool-calls-and-never-leaves-the-machine-vihln3unw) · XDA Developers · 1 upvotes · 0 comments - [Muse-Glimmer-30B and Qwen 3.8-27B models now available on Amazon SageMaker JumpStart](https://daily.dev/posts/muse-glimmer-30b-and-qwen-3-8-27b-models-now-available-on-amazon-sagemaker-jumpstart-tdhrrjsjl) · AWS · 3 upvotes · 1 comments - [Fine-tune Meta AI's Muse Glimmer 100% locally](https://daily.dev/posts/fine-tune-meta-ai-s-muse-glimmer-100-locally-nvrlxatlc) · Daily Dose of Data Science \| Avi Chawla \| Substack · 0 upvotes · 0 comments --- Tags: [#ai-inference](https://daily.dev/tags/ai-inference), [#deep-learning](https://daily.dev/tags/deep-learning), [#llm](https://daily.dev/tags/llm) [View this post on daily.dev](https://daily.dev/posts/muse-glimmer-30b-architecture-notes-dkdmm5p34) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Muse Glimmer 30B Architecture Notes","url":"https://daily.dev/posts/muse-glimmer-30b-architecture-notes-dkdmm5p34","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/muse-glimmer-30b-architecture-notes-dkdmm5p34"},"datePublished":"2026-08-11T20:53:17.735Z","dateModified":"2026-10-14T02:28:16.410Z","description":"Meta released Muse Glimmer 30B, a 30B multimodal reasoning model with a Gemma-like architecture. Key architectural highlights include a 131k context window,...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/e511fbbbcfa52da9dde18b9aa6e63dd5?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/e511fbbbcfa52da9dde18b9aa6e63dd5?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"Sebastian Raschka","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Sebastian Raschka","logo":"https://media.daily.dev/image/upload/t_logo,f_auto/v1/logos/eb22d2e73d074c5598d31588ad16a51c","url":"https://daily.dev/sources/sebastianraschka"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/muse-glimmer-30b-architecture-notes-dkdmm5p34","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"ai-inference,deep-learning,llm","timeRequired":"PT3M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Sebastian Raschka","item":"https://daily.dev/sources/sebastianraschka"},{"@type":"ListItem","position":3,"name":"Muse Glimmer 30B Architecture Notes"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/muse-glimmer-30b-architecture-notes-dkdmm5p34#faq","mainEntity":[{"@type":"Question","name":"What is the KV cache size per token for Meta Muse Glimmer 30B compared to Qwen3.6 27B and Gemma 4 31B?","acceptedAnswer":{"@type":"Answer","text":"Muse Glimmer 30B has a KV cache of 52 KiB per token in BF16, compared to 64 KiB for Qwen3.6 27B and 840 KiB for Gemma 4 31B. This extreme efficiency comes from its 32 query heads to just 2 KV heads GQA ratio, making it particularly memory-efficient for long-context and agentic workloads. Engineers choosing models for memory-constrained agentic deployments track comparisons like this on daily.dev."}},{"@type":"Question","name":"What is the SWA to GQA ratio in Meta Muse Glimmer 30B and how does it compare to Gemma 4?","acceptedAnswer":{"@type":"Answer","text":"Muse Glimmer 30B uses a 3:1 local-to-global (SWA:GQA) ratio in its hybrid attention design. Gemma 4 uses a 5:1 ratio for comparison. Glimmer also applies gated attention — a sigmoid gate on the attention output — to both its SWA and GQA layers, which has become increasingly common in recent model releases. Researchers tracking LLM architectural trends across model releases find those discussions on daily.dev."}}]} ```