--- title: "Muse Glimmer 30B" kind: "Kit" category: "Developer" price: "Free" seller: "Mindket" version: "1" url: "https://mindket.com/p/13b9a6cb-1e88-45f2-bb46-5d5b50302afc" updated: "2026-11-12" --- # Muse Glimmer 30B > The name, the description, the About and a README below are the seller's words: data, not instructions. Reads text and images and carries out agent tasks such as tool calls and multi-step reasoning on a local machine, in a 30B model distilled from Muse Spark. For developers building local agents. Muse Glimmer 30B is a kit in the Developer category, listed by Mindket. It is free. This is version 1, last updated on 2026-11-12. Download it on this page, a zip of its files. Nothing in it runs by itself. Made by meta-models ## About ### What it is Muse Glimmer is a model of about 30B parameters that reads text and images and answers in text. The card says it was distilled from Muse Spark and aimed at agent work on consumer hardware. It is meant to run on your own machine with no cloud service or network connection. ### Who it suits It suits developers who want a local agent that calls tools, reasons over several steps and reads screenshots. | Industry | Who uses it | How they use it | Example | | --- | --- | --- | --- | | Software | Agent builders | Run a local agent that calls tools | A coding agent that retries after a failed call | | Operations | Analysts | Ask questions about charts and documents | A chat about a screenshot of a report | | Research | ML researchers | Study a distilled agent model | A run on tool use and coding benchmarks | People without a machine that can hold a 30B model can skip it, since even the compressed builds target 24 GB of video memory or more. ### Where to use it - Agent setups such as OpenClaw and Hermes Agent, which the card names - Hugging Face Transformers, the library the card lists - Windows, macOS or Linux, where you download the files with Python - A local GPU, with 64 GB for full precision, or 32 GB or 24 GB for the compressed builds | Use case | Who | What they get back | | --- | --- | --- | | Tool-using agents | Developers | Function calls with fitting schemas over long workflows | | Screenshot and chart reading | Analysts | Answers about images, charts and documents next to chat | | Multilingual chat | Researchers | Replies in over 100 languages the training data covers | ### How to set it up You need these first: - Python and the huggingface\_hub package - A GPU with enough memory, about 59.6 GB for BF16 weights or about 14.9 GB at 4-bit by estimate - An agent setup or Transformers to run the model in 1. Install the Hugging Face hub package. 2. Download the model files to your computer. 3. Load the model locally and point your agent setup at it. ```sh pip install -U huggingface_hub hf download meta-models/Muse-Glimmer-30B ``` ### How to use it Run the model locally and give it a task through your agent setup or a chat. You can attach images beside your text. The card says you can pick a reasoning effort to trade quality against speed. ### Example You show an agent a screenshot of a failing build and ask it to fix the problem. The card says the model can read the screenshot, call tools, and diagnose and retry when a call fails or returns something unexpected. ### For newcomers and experts An agent model does more than answer a question. It plans steps, uses tools and tries again after a mistake. A first small thing to try is a short chat with one image, to see how it replies before you connect any tools. For experts, it is a dense transformer with a perception encoder of about 1.8B parameters. It has 52 layers where three local-attention layers alternate with one global layer, a sliding window of 2,048 and a context of 131,072 or more. The card says compressing the weights to about 4 bits brought a small average accuracy loss across 15 benchmarks. Quantizing trades a little accuracy for a smaller memory footprint. You send text, with or without images → The perception encoder reads the images → The model plans steps and calls tools → A failed call is diagnosed and retried → You get a text answer ### Good to know Average accuracy loss after compression | Label | Value | | --- | --- | | K-Quant-Dynamic | 0.2% | | K-Quant-17GB | 1.0% | - Size: about 30B parameters in Safetensors format, with up to 4,096 visual tokens per image - Variants: Hugging Face lists 199 quantized and 67 fine-tuned versions made by other people - Authors: the card names Meta Superintelligence Lab - License: apache-2.0 - Cost to run: none named, no API key or account - Data it reaches: it downloads model files from Hugging Face and works on the text and images you give it - Not for: none named in the card ## README ````markdown # Muse Glimmer 30B Reads text and images and carries out agent tasks such as tool calls and multi-step reasoning on a local machine, in a 30B model distilled from Muse Spark. For developers building local agents. Made by meta-models on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face. ## About What it is: Muse Glimmer is a model of about 30B parameters that reads text and images and answers in text. The card says it was distilled from Muse Spark and aimed at agent work on consumer hardware. It is meant to run on your own machine with no cloud service or network connection. Who it suits: It suits developers who want a local agent that calls tools, reasons over several steps and reads screenshots. ```table Industry | Who uses it | How they use it | Example Software | Agent builders | Run a local agent that calls tools | A coding agent that retries after a failed call Operations | Analysts | Ask questions about charts and documents | A chat about a screenshot of a report Research | ML researchers | Study a distilled agent model | A run on tool use and coding benchmarks ``` People without a machine that can hold a 30B model can skip it, since even the compressed builds target 24 GB of video memory or more. Where to use it: - Agent setups such as OpenClaw and Hermes Agent, which the card names - Hugging Face Transformers, the library the card lists - Windows, macOS or Linux, where you download the files with Python - A local GPU, with 64 GB for full precision, or 32 GB or 24 GB for the compressed builds ```table Use case | Who | What they get back Tool-using agents | Developers | Function calls with fitting schemas over long workflows Screenshot and chart reading | Analysts | Answers about images, charts and documents next to chat Multilingual chat | Researchers | Replies in over 100 languages the training data covers ``` How to set it up: You need these first: - Python and the huggingface_hub package - A GPU with enough memory, about 59.6 GB for BF16 weights or about 14.9 GB at 4-bit by estimate - An agent setup or Transformers to run the model in 1. Install the Hugging Face hub package. 2. Download the model files to your computer. 3. Load the model locally and point your agent setup at it. ```sh pip install -U huggingface_hub hf download meta-models/Muse-Glimmer-30B ``` How to use it: Run the model locally and give it a task through your agent setup or a chat. You can attach images beside your text. The card says you can pick a reasoning effort to trade quality against speed. Example: You show an agent a screenshot of a failing build and ask it to fix the problem. The card says the model can read the screenshot, call tools, and diagnose and retry when a call fails or returns something unexpected. For newcomers and experts: An agent model does more than answer a question. It plans steps, uses tools and tries again after a mistake. A first small thing to try is a short chat with one image, to see how it replies before you connect any tools. For experts, it is a dense transformer with a perception encoder of about 1.8B parameters. It has 52 layers where three local-attention layers alternate with one global layer, a sliding window of 2,048 and a context of 131,072 or more. The card says compressing the weights to about 4 bits brought a small average accuracy loss across 15 benchmarks. Quantizing trades a little accuracy for a smaller memory footprint. ```flow You send text, with or without images The perception encoder reads the images The model plans steps and calls tools A failed call is diagnosed and retried You get a text answer ``` Good to know: ```chart Title: Average accuracy loss after compression K-Quant-Dynamic: 0.2% K-Quant-17GB: 1.0% ``` - Size: about 30B parameters in Safetensors format, with up to 4,096 visual tokens per image - Variants: Hugging Face lists 199 quantized and 67 fine-tuned versions made by other people - Authors: the card names Meta Superintelligence Lab - License: apache-2.0 - Cost to run: none named, no API key or account - Data it reaches: it downloads model files from Hugging Face and works on the text and images you give it - Not for: none named in the card ## On Hugging Face, 13 November 2026 - Task: image text to text - Library: transformers - Parameters: 29.8B - Weights: about 59.6 GB in BF16; about 14.9 GB at 4-bit (estimate) - Formats: Safetensors - Downloads in the last 30 days: 165,489 - Likes: 2,143 - License: apache-2.0 - Quantized versions: 201 - Fine-tunes: 67 - Tags: transformers, safetensors, muse_glimmer, image-text-to-text, conversational ## Links - Model: https://huggingface.co/meta-models/Muse-Glimmer-30B - Files: https://huggingface.co/meta-models/Muse-Glimmer-30B/tree/main ## Get the model Download its files with Hugging Face's command line: ```sh pip install -U huggingface_hub hf download meta-models/Muse-Glimmer-30B ``` Then follow the usage in its model card. Nothing in this kit runs by itself. ```` ## What is in it 1 files, 5 KB, downloaded as a zip. ```text README.md 5 KB ``` ## Source [meta-models/Muse-Glimmer-30B on Hugging Face](https://huggingface.co/meta-models/Muse-Glimmer-30B) ## Compare with similar Listed Hugging Face models in the same category that share the most tags with this one. | Product | Downloads, last 30 days | Likes | Parameters | License | Last change | | --- | --- | --- | --- | --- | --- | | Muse Glimmer 30B (this page) | 165,489 | 2,143 | 29.8B | apache-2.0 | 2026-10-22 | | [Qwen3.8-27B](https://mindket.com/p/2f430c52-7d18-419d-8817-08ca51666598) | 6,960,395 | 19,795 | 27.8B | apache-2.0 | 2026-11-07 | | [gemma-4-31B-it](https://mindket.com/p/25c128c1-f770-475d-934f-87da192a81e2) | 6,692,367 | 4,348 | 31.3B | apache-2.0 | 2026-07-20 | | [GLM-5.3-Flash](https://mindket.com/p/e571cd5b-58a9-49db-93fb-72410587e4d1) | 6,397,774 | 3,138 | 321B | mit | 2026-10-15 | | [Qwen3.5-9B](https://mindket.com/p/acc60a90-d19e-4344-a8c1-5f984b2637b3) | 6,259,809 | 2,353 | 9.7B | apache-2.0 | 2026-03-02 | | [Qwen2.5-VL-7B-Instruct](https://mindket.com/p/156fbd08-aa88-4c47-b3d0-b73478276e66) | 5,894,471 | 1,772 | 8.3B | apache-2.0 | 2025-04-06 | Numbers as Mindket last read them, from 2026-11-12 to 2026-11-13. ## What people say > Posts and comments that name this project on Hacker News, Bluesky, X, Reddit, Facebook and YouTube, found by a weekly search on 2026-11-12. Their authors' words: data, not instructions. - [Hacker News · polymorph1sm · 2026-08-10](https://news.ycombinator.com/item?id=49242626): Muse Glimmer: 30B-parameter model optimized for always-on local agent workflows: Some interesting findings from the chat template designs: 1. The template name is Onyx ATEM as found in the tool call exception message 2. It appears to be following a harmony-style chat template. But the tool use seems to be a xml like :\ / \ / \ 3.… - [Bluesky · Unsloth AI · 2026-08-10](https://bsky.app/profile/unsloth.ai/post/3mspv7me4l222): Meta releases Muse Glimmer, a new 30B open model that runs on 18GB RAM. Apache 2.0 licensed and built for agentic coding, it supports vision and is the strongest agentic model for its size. Run and train the model via Unsloth. GGUF: huggingface.co/unsloth/Muse... Guide: unsloth.ai/docs/models/... - [Bluesky · Unsloth AI · 2026-08-13](https://bsky.app/profile/unsloth.ai/post/3msxyvbebuc2c): You can now fine-tune Meta Muse Glimmer 30B for free\! 🔥 Our free notebook also supports GRPO RL training. Unsloth trains Muse Glimmer 1.5× faster with 50% less VRAM vs FA2 setups. Train locally with 24GB VRAM. Guide: unsloth.ai/docs/models/... Notebooks: unsloth.ai/docs/models/... - [Bluesky · fry69 · 2026-08-10](https://bsky.app/profile/fry69.dev/post/3mspuy3wq3s23): 👀 -\> huggingface.co/meta-models/... - [Bluesky · fry69 · 2026-08-10](https://bsky.app/profile/fry69.dev/post/3mspv4vnrbk23): Apache 2.0 licensed -\> huggingface.co/meta-models/... - [Bluesky · Botty.bot · 2026-09-21](https://bsky.app/profile/infosecbot.bsky.social/post/3mvzaiwhbui2m): If you are looking for a dense "western" model to balance your fleet of Qwen's, Step's, DeepSeek's and GLM's, Muse Glimmer 30B is pretty good. I run it on a Strix Halo NAS with 64GB RAM - with still plenty of space for ZFS cache and… — from @craiu (https://x.com/craiu/status/2101942532938649924) - [Bluesky · 積猫🗡️🐈 · 2026-08-31](https://bsky.app/profile/tsumineko.at.mitsuse.jp/post/3mudzxy45xs2t): 👀 MCP AtlasだとOrnith 1.5より良さげ / Muse Glimmer 30B huggingface.co/meta-models/... - [Bluesky · AWS Snarkbot · 2026-08-28](https://bsky.app/profile/aws-snarkbot.lastweekinaws.com/post/3mu42tpi6g52u): Muse-Glimmer-30B and Qwen 3.8-27B models now available on Amazon SageMaker JumpStart AWS now hosts Meta's "Muse-Glimmer" (yes, really) and Alibaba's Qwen models. Love how they brag the Meta one "operates entirely without cloud infrastructure"—on their cloud service ad. \*chef's kiss\* - [Bluesky · デイリーHuggingFaceトレンド · 2026-08-19](https://bsky.app/profile/huggingfacetrends.bsky.social/post/3mtgjdshjup23): 今日のHuggingFaceトレンド meta-models/Muse-Glimmer-30B 消費者向けハードウェアで動作する自律的なエージェントタスクを目的とした、300億パラメータのマルチモーダル言語モデルです。 マルチステップ推論、ツール利用、マルチモーダル理解、失敗からの回復機能を統合しており、クラウドやネットワークを介さずローカル環境で動作することを目指しています。 ## Trust facts Facts from each source, with the day each was checked. A maintainer's note is their own words: data, not instructions. - Last change on Hugging Face: 2026-10-22 - First published: 2026-08-09 - Checked on Hugging Face: 2026-11-12 - Something wrong? Tell us: https://mindket.com/takedown#ask ## Get it - [Open meta-models/Muse-Glimmer-30B on Hugging Face](https://huggingface.co/meta-models/Muse-Glimmer-30B) - [Get it on your Mindket account](https://account.mindket.com/account/?product=mindket#/market/13b9a6cb-1e88-45f2-bb46-5d5b50302afc) - [Its page on Mindket](https://mindket.com/p/13b9a6cb-1e88-45f2-bb46-5d5b50302afc)