--- title: Muse Glimmer 30B runs on my 32GB GPU, handles its own... description: Meta released Muse Glimmer 30B, an Apache 2.0 dense model built specifically for agentic tool-calling loops rather than chat, and the author put it through... canonical: https://daily.dev/posts/muse-glimmer-30b-runs-on-my-32gb-gpu-handles-its-own-tool-calls-and-never-leaves-the-machine-vihln3unw twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Muse Glimmer 30B runs on my 32GB GPU, handles its own tool calls, and never leaves the machine | daily.dev og:description: Meta released Muse Glimmer 30B, an Apache 2.0 dense model built specifically for agentic tool-calling loops rather than chat, and the author put it through... og:url: https://daily.dev/posts/muse-glimmer-30b-runs-on-my-32gb-gpu-handles-its-own-tool-calls-and-never-leaves-the-machine-vihln3unw og:image: https://api.daily.dev/og/posts/vihlN3UNw.png og:image:alt: Muse Glimmer 30B runs on my 32GB GPU, handles its own tool calls, and never leaves the machine 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 runs on my 32GB GPU, handles its own tool calls, and never leaves the machine **[XDA Developers](https://daily.dev/sources/xda-developers)** · 5 min read · 1 upvotes · 0 comments ## Summary Meta released Muse Glimmer 30B, an Apache 2.0 dense model built specifically for agentic tool-calling loops rather than chat, and the author put it through real-world testing on an RTX 5090. The model uses a lean KV cache architecture (32 query heads, 2 KV heads, 3:1 sliding-window attention) that keeps a full 131K context on a single 32GB card, hitting 74.9 tokens/sec baseline and 233 tokens/sec with DFlash draft decoding via AMD's Lemonade runtime. Testing against seven adversarial tool-calling scenarios, including a prompt injection attempt to shut off rack power, showed the model correctly diagnosing failures, recovering from a dead DNS tool by reading a hosts file, and explicitly refusing an injected instruction while explaining why. The author frames it as the first local 30B model trustworthy enough to run unattended for tasks involving credentials, private documents, or NDA-bound client code. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### What is Muse Glimmer 30B and how is it different from other local LLMs? Muse Glimmer is Meta's dense 29.6B-parameter transformer released under Apache 2.0, paired with a 1.8B perception encoder for images, designed specifically for agentic tool-calling rather than chat. It uses 32 query heads with only 2 KV heads and a 3:1 sliding-window to global attention ratio, giving a KV cache of about 52 KiB per token versus 64 KiB for Qwen 3.6 27B and 840 KiB for Gemma 4 31B, letting a full 131K token context fit on a single 32GB GPU. _Developers weighing local versus cloud agents for sensitive workloads can track model releases like this on daily.dev._ ### How fast does Muse Glimmer 30B run on an RTX 5090 with Lemonade? Meta reports 74.9 tokens per second baseline on an RTX 5090, rising to 233.4 tokens per second with DFlash draft decoding enabled, a 3.1x speedup from a drafter that guesses 16 tokens ahead. In real agent-loop testing with reasoning and tool-call JSON included, throughput averaged about 110 tokens per second end-to-end with the full 131,072-token context loaded, using 27.6GB of the card's 31.5GB VRAM. _Anyone benchmarking local GPU inference speeds can follow real-world numbers like these on daily.dev._ ### Can local LLM agents recover from failed tool calls or resist prompt injection? Yes, Muse Glimmer 30B was tested against seven adversarial scenarios including a malformed API schema, a 503 error, a dead DNS tool, and a text file containing an injected instruction to power off a rack switch. It resolved the dead DNS issue by reading a hosts file directly, and on the injection attempt it declined and explained that the text was file content, not a user instruction, rather than blindly executing it. _Teams building tool-calling agents can watch how local models on daily.dev handle failure recovery and injection resistance._ --- Tags: [#llm](https://daily.dev/tags/llm), [#ai-agents](https://daily.dev/tags/ai-agents), [#local-ai](https://daily.dev/tags/local-ai) [View this post on daily.dev](https://daily.dev/posts/muse-glimmer-30b-runs-on-my-32gb-gpu-handles-its-own-tool-calls-and-never-leaves-the-machine-vihln3unw) ```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 runs on my 32GB GPU, handles its own tool calls, and never leaves the machine","url":"https://daily.dev/posts/muse-glimmer-30b-runs-on-my-32gb-gpu-handles-its-own-tool-calls-and-never-leaves-the-machine-vihln3unw","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/muse-glimmer-30b-runs-on-my-32gb-gpu-handles-its-own-tool-calls-and-never-leaves-the-machine-vihln3unw"},"datePublished":"2026-08-30T23:04:52.250Z","dateModified":"2026-08-30T23:10:08.071Z","description":"Meta released Muse Glimmer 30B, an Apache 2.0 dense model built specifically for agentic tool-calling loops rather than chat, and the author put it through...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/6808b578cac9b2b0979e0dfe178d02ac?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/6808b578cac9b2b0979e0dfe178d02ac?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"XDA Developers","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":"XDA Developers","logo":"https://media.daily.dev/image/upload/s--kCg6yyAP--/f_auto,q_auto/v1774964407/logos/xda-developers?_a=BAMAMiWQ0","url":"https://daily.dev/sources/xda-developers"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/muse-glimmer-30b-runs-on-my-32gb-gpu-handles-its-own-tool-calls-and-never-leaves-the-machine-vihln3unw","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":1},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"llm,ai-agents,local-ai","timeRequired":"PT5M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"XDA Developers","item":"https://daily.dev/sources/xda-developers"},{"@type":"ListItem","position":3,"name":"Muse Glimmer 30B runs on my 32GB GPU, handles its own tool calls, and never leaves the machine"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/muse-glimmer-30b-runs-on-my-32gb-gpu-handles-its-own-tool-calls-and-never-leaves-the-machine-vihln3unw#faq","mainEntity":[{"@type":"Question","name":"What is Muse Glimmer 30B and how is it different from other local LLMs?","acceptedAnswer":{"@type":"Answer","text":"Muse Glimmer is Meta's dense 29.6B-parameter transformer released under Apache 2.0, paired with a 1.8B perception encoder for images, designed specifically for agentic tool-calling rather than chat. It uses 32 query heads with only 2 KV heads and a 3:1 sliding-window to global attention ratio, giving a KV cache of about 52 KiB per token versus 64 KiB for Qwen 3.6 27B and 840 KiB for Gemma 4 31B, letting a full 131K token context fit on a single 32GB GPU. Developers weighing local versus cloud agents for sensitive workloads can track model releases like this on daily.dev."}},{"@type":"Question","name":"How fast does Muse Glimmer 30B run on an RTX 5090 with Lemonade?","acceptedAnswer":{"@type":"Answer","text":"Meta reports 74.9 tokens per second baseline on an RTX 5090, rising to 233.4 tokens per second with DFlash draft decoding enabled, a 3.1x speedup from a drafter that guesses 16 tokens ahead. In real agent-loop testing with reasoning and tool-call JSON included, throughput averaged about 110 tokens per second end-to-end with the full 131,072-token context loaded, using 27.6GB of the card's 31.5GB VRAM. Anyone benchmarking local GPU inference speeds can follow real-world numbers like these on daily.dev."}},{"@type":"Question","name":"Can local LLM agents recover from failed tool calls or resist prompt injection?","acceptedAnswer":{"@type":"Answer","text":"Yes, Muse Glimmer 30B was tested against seven adversarial scenarios including a malformed API schema, a 503 error, a dead DNS tool, and a text file containing an injected instruction to power off a rack switch. It resolved the dead DNS issue by reading a hosts file directly, and on the injection attempt it declined and explained that the text was file content, not a user instruction, rather than blindly executing it. Teams building tool-calling agents can watch how local models on daily.dev handle failure recovery and injection resistance."}}]} ```