{"id":10707,"date":"2026-11-17T10:59:39","date_gmt":"2026-11-17T01:59:39","guid":{"rendered":"https:\/\/localmodelwatch.tsuchitsuchi.com\/?p=10707"},"modified":"2026-11-17T11:09:04","modified_gmt":"2026-11-17T02:09:04","slug":"muse-glimmer-30b-gguf-2","status":"publish","type":"post","link":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/11\/17\/muse-glimmer-30b-gguf-2\/","title":{"rendered":"Muse-Glimmer-30B-GGUF 30B Multimodal Local Agent Model: 24GB+ VRAM"},"content":{"rendered":"<p><!-- lmw:facts --><\/p>\n<h2>At a Glance<\/h2>\n<div class=\"lmw-table-scroll\" tabindex=\"0\" style=\"overflow-x:auto;-webkit-overflow-scrolling:touch;max-width:100%;\">\n<table style=\"width:max-content;min-width:100%;border-collapse:collapse;\">\n<thead>\n<tr>\n<th>Item<\/th>\n<th>Value<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>Repository<\/td>\n<td><a href=\"https:\/\/huggingface.co\/ggml-org\/Muse-Glimmer-30B-GGUF\">ggml-org\/Muse-Glimmer-30B-GGUF<\/a><\/td>\n<\/tr>\n<tr>\n<td>Published<\/td>\n<td>2026-10-08<\/td>\n<\/tr>\n<tr>\n<td>License<\/td>\n<td>apache-2.0<\/td>\n<\/tr>\n<tr>\n<td>Formats<\/td>\n<td><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/format-gguf-en\/\">GGUF<\/a><\/td>\n<\/tr>\n<tr>\n<td>Source type<\/td>\n<td>Primary source (the publisher itself)<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n<p><em>Values determined by this site&#8217;s code when the information was collected. Dates are JST.<\/em><\/p>\n<p><!-- \/lmw:facts --><\/p>\n<h2>Overview<\/h2>\n<p>ggml-org has released the GGUF quantized version &#8220;ggml-org\/Muse-Glimmer-30B-GGUF&#8221; of &#8220;Muse-Glimmer-30B&#8221;, a 30B-class multimodal local agent model developed by Meta Superintelligence Lab. The original model, &#8220;Muse-Glimmer-30B&#8221;, is a causal language model with approximately 29.6B parameters, distilled from the higher-level model &#8220;Muse Spark&#8221; and specifically designed to execute autonomous agent tasks on consumer-grade hardware. This model integrates multi-step reasoning, reliable tool use, multimodal comprehension, and self-correction during errors (recovery from failure) into a single locally runnable model.<\/p>\n<h2>Specifications<\/h2>\n<p>The specifications of the original model &#8220;<a href=\"https:\/\/huggingface.co\/meta-models\/Muse-Glimmer-30B\">meta-models\/Muse-Glimmer-30B<\/a>&#8221; are as follows:<\/p>\n<ul>\n<li><strong>Total Parameters<\/strong>: Approx. 29.6B (including vision encoder)<\/li>\n<li><strong>Architecture<\/strong>: Dense Causal Transformer with Perception Encoder<\/li>\n<li><strong>Context Length<\/strong>: 131,072+<\/li>\n<li><strong>Number of Layers<\/strong>: 52<\/li>\n<li><strong>Attention Pattern<\/strong>: Repetition of [Local, Local, Local, Global]<\/li>\n<li><strong>Sliding Window Size<\/strong>: 2048<\/li>\n<li><strong>Attention Heads (Q \/ KV)<\/strong>: 32 \/ 2 (GQA ratio 16:1)<\/li>\n<li><strong>Perception Encoder<\/strong>: Approx. 1.8B parameter ViT-G\/14 (50 layers, width 1536, patch size 14)<\/li>\n<li><strong>Vocabulary Size<\/strong>: 202,048<\/li>\n<li><strong>Max Visual Tokens per Image<\/strong>: 4,096<\/li>\n<li><strong>Supported Modalities<\/strong>: Input: Text + Image, Output: Text<\/li>\n<li><strong>Knowledge Cutoff<\/strong>: January 4, 2026<\/li>\n<\/ul>\n<h2>Performance<\/h2>\n<p>The performance information shown below consists entirely of measurements by the developers stated in the public card of the original model &#8220;<a href=\"https:\/\/huggingface.co\/meta-models\/Muse-Glimmer-30B\">meta-models\/Muse-Glimmer-30B<\/a>&#8220;. Actual performance may vary slightly due to quantization or conversion to GGUF.<\/p>\n<h3>Benchmark Evaluation Comparison<\/h3>\n<p>In the developers&#8217; evaluation, Muse Glimmer-30B is compared with leading models of the same scale class (Gemma4-31B, Qwen3.6-27B).<\/p>\n<div class=\"lmw-table-scroll\" tabindex=\"0\" style=\"overflow-x:auto;-webkit-overflow-scrolling:touch;max-width:100%;\">\n<p class=\"lmw-table-hint\" style=\"margin:0 0 4px;font-size:0.85em;opacity:0.7;\">\u2192 Scroll horizontally to see all columns<\/p>\n<table style=\"width:max-content;min-width:100%;border-collapse:collapse;\">\n<thead>\n<tr>\n<th style=\"text-align: left;\">Category<\/th>\n<th style=\"text-align: left;\">Benchmark<\/th>\n<th style=\"text-align: center;\">Muse Glimmer-30B High Reasoning<\/th>\n<th style=\"text-align: center;\">Gemma4-31B Thinking Mode<\/th>\n<th style=\"text-align: center;\">Qwen3.6-27B Thinking Mode<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td style=\"text-align: left;\"><em>General Agentic<\/em><\/td>\n<td style=\"text-align: left;\">MCP Atlas (Public)<\/td>\n<td style=\"text-align: center;\"><strong>75.5<\/strong><\/td>\n<td style=\"text-align: center;\">54.2<\/td>\n<td style=\"text-align: center;\">62.5<\/td>\n<\/tr>\n<tr>\n<td style=\"text-align: left;\"><\/td>\n<td style=\"text-align: left;\">DeepSearch QA<\/td>\n<td style=\"text-align: center;\"><strong>74.6<\/strong><\/td>\n<td style=\"text-align: center;\">61.7<\/td>\n<td style=\"text-align: center;\">71.1<\/td>\n<\/tr>\n<tr>\n<td style=\"text-align: left;\"><\/td>\n<td style=\"text-align: left;\">\ud835\uded53-Banking<\/td>\n<td style=\"text-align: center;\"><strong>23.5<\/strong><\/td>\n<td style=\"text-align: center;\">15.1<\/td>\n<td style=\"text-align: center;\">16.7<\/td>\n<\/tr>\n<tr>\n<td style=\"text-align: left;\"><\/td>\n<td style=\"text-align: left;\">WildClawBench<\/td>\n<td style=\"text-align: center;\"><strong>47.6<\/strong><\/td>\n<td style=\"text-align: center;\">37.6<\/td>\n<td style=\"text-align: center;\">43.2<\/td>\n<\/tr>\n<tr>\n<td style=\"text-align: left;\"><\/td>\n<td style=\"text-align: left;\">GDPVal-AA v2<\/td>\n<td style=\"text-align: center;\">953<\/td>\n<td style=\"text-align: center;\">811<\/td>\n<td style=\"text-align: center;\"><strong>1141<\/strong><\/td>\n<\/tr>\n<tr>\n<td style=\"text-align: left;\"><\/td>\n<td style=\"text-align: left;\">Gaia2<\/td>\n<td style=\"text-align: center;\"><strong>43.3<\/strong><\/td>\n<td style=\"text-align: center;\">36.4<\/td>\n<td style=\"text-align: center;\">40.0<\/td>\n<\/tr>\n<tr>\n<td style=\"text-align: left;\"><\/td>\n<td style=\"text-align: left;\">SkillsBench (with skills)<\/td>\n<td style=\"text-align: center;\">44.3<\/td>\n<td style=\"text-align: center;\">32.4<\/td>\n<td style=\"text-align: center;\"><strong>46.6<\/strong><\/td>\n<\/tr>\n<tr>\n<td style=\"text-align: left;\"><\/td>\n<td style=\"text-align: left;\">OSWorld-Verified<\/td>\n<td style=\"text-align: center;\">65.9<\/td>\n<td style=\"text-align: center;\">58.5<\/td>\n<td style=\"text-align: center;\"><strong>75.6<\/strong><\/td>\n<\/tr>\n<tr>\n<td style=\"text-align: left;\"><em>Agentic Coding<\/em><\/td>\n<td style=\"text-align: left;\">SWE-Bench Pro<\/td>\n<td style=\"text-align: center;\"><strong>51.2<\/strong><\/td>\n<td style=\"text-align: center;\">36.9<\/td>\n<td style=\"text-align: center;\">50.2<\/td>\n<\/tr>\n<tr>\n<td style=\"text-align: left;\"><\/td>\n<td style=\"text-align: left;\">SWE-Bench Verified<\/td>\n<td style=\"text-align: center;\">76.0<\/td>\n<td style=\"text-align: center;\">66.6<\/td>\n<td style=\"text-align: center;\"><strong>77.2<\/strong><\/td>\n<\/tr>\n<tr>\n<td style=\"text-align: left;\"><\/td>\n<td style=\"text-align: left;\">TerminalBench 2.1 (with terminus2)<\/td>\n<td style=\"text-align: center;\">51.7<\/td>\n<td style=\"text-align: center;\">43.4<\/td>\n<td style=\"text-align: center;\"><strong>60.7<\/strong><\/td>\n<\/tr>\n<tr>\n<td style=\"text-align: left;\"><\/td>\n<td style=\"text-align: left;\">SciCode<\/td>\n<td style=\"text-align: center;\"><strong>43.6<\/strong><\/td>\n<td style=\"text-align: center;\">43.4<\/td>\n<td style=\"text-align: center;\">39.8<\/td>\n<\/tr>\n<tr>\n<td style=\"text-align: left;\"><em>Multimodal<\/em><\/td>\n<td style=\"text-align: left;\">Charxiv Reasoning<\/td>\n<td style=\"text-align: center;\"><strong>78.8<\/strong><\/td>\n<td style=\"text-align: center;\">77.7<\/td>\n<td style=\"text-align: center;\">78.4<\/td>\n<\/tr>\n<tr>\n<td style=\"text-align: left;\"><\/td>\n<td style=\"text-align: left;\">ScreenSpot Pro<\/td>\n<td style=\"text-align: center;\">75.4<\/td>\n<td style=\"text-align: center;\">75.9<\/td>\n<td style=\"text-align: center;\"><strong>76.1<\/strong><\/td>\n<\/tr>\n<tr>\n<td style=\"text-align: left;\"><\/td>\n<td style=\"text-align: left;\">OmniDocBench v1.5<\/td>\n<td style=\"text-align: center;\">75.8<\/td>\n<td style=\"text-align: center;\">72.5<\/td>\n<td style=\"text-align: center;\"><strong>77.8<\/strong><\/td>\n<\/tr>\n<tr>\n<td style=\"text-align: left;\"><\/td>\n<td style=\"text-align: left;\">MMMU Pro<\/td>\n<td style=\"text-align: center;\">74<\/td>\n<td style=\"text-align: center;\">73<\/td>\n<td style=\"text-align: center;\"><strong>75<\/strong><\/td>\n<\/tr>\n<tr>\n<td style=\"text-align: left;\"><em>General Capabilities and Reasoning<\/em><\/td>\n<td style=\"text-align: left;\">IFBench<\/td>\n<td style=\"text-align: center;\"><strong>77.0<\/strong><\/td>\n<td style=\"text-align: center;\">76.0<\/td>\n<td style=\"text-align: center;\">70.8<\/td>\n<\/tr>\n<tr>\n<td style=\"text-align: left;\"><\/td>\n<td style=\"text-align: left;\">AIME 2026<\/td>\n<td style=\"text-align: center;\"><strong>94.7<\/strong><\/td>\n<td style=\"text-align: center;\">89.2<\/td>\n<td style=\"text-align: center;\">94.1<\/td>\n<\/tr>\n<tr>\n<td style=\"text-align: left;\"><\/td>\n<td style=\"text-align: left;\">GPQA Diamond (AA)<\/td>\n<td style=\"text-align: center;\">83.5<\/td>\n<td style=\"text-align: center;\"><strong>85.7<\/strong><\/td>\n<td style=\"text-align: center;\">84.2<\/td>\n<\/tr>\n<tr>\n<td style=\"text-align: left;\"><\/td>\n<td style=\"text-align: left;\">HLE Text (AA)<\/td>\n<td style=\"text-align: center;\">22.0<\/td>\n<td style=\"text-align: center;\"><strong>23.6<\/strong><\/td>\n<td style=\"text-align: center;\">23.1<\/td>\n<\/tr>\n<tr>\n<td style=\"text-align: left;\"><\/td>\n<td style=\"text-align: left;\">AA-LCR<\/td>\n<td style=\"text-align: center;\"><strong>80.0<\/strong><\/td>\n<td style=\"text-align: center;\">68.3<\/td>\n<td style=\"text-align: center;\">73.3<\/td>\n<\/tr>\n<tr>\n<td style=\"text-align: left;\"><\/td>\n<td style=\"text-align: left;\">Beam128K<\/td>\n<td style=\"text-align: center;\"><strong>65.1<\/strong><\/td>\n<td style=\"text-align: center;\">58.2<\/td>\n<td style=\"text-align: center;\">63.0<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n<p>These evaluation results show that Muse Glimmer-30B achieves scores significantly higher than other models in many &#8220;General Agentic&#8221; benchmarks (such as 75.5 on MCP Atlas and 74.6 on DeepSearch QA). In &#8220;Agentic Coding&#8221; as well, it demonstrates high aptitude, recording 51.2 on SWE-Bench Pro. On the other hand, in validation tasks such as OSWorld-Verified (65.9) and TerminalBench 2.1 (51.7), or specific reasoning benchmarks like GPQA Diamond (83.5), it falls slightly behind Qwen3.6-27B and Gemma4-31B.<\/p>\n<h3>Scientific and Safety Evaluation (Preparedness)<\/h3>\n<p>The comparison results in fields measuring scientific knowledge and laboratory debugging capabilities are as follows:<\/p>\n<div class=\"lmw-table-scroll\" tabindex=\"0\" style=\"overflow-x:auto;-webkit-overflow-scrolling:touch;max-width:100%;\">\n<p class=\"lmw-table-hint\" style=\"margin:0 0 4px;font-size:0.85em;opacity:0.7;\">\u2192 Scroll horizontally to see all columns<\/p>\n<table style=\"width:max-content;min-width:100%;border-collapse:collapse;\">\n<thead>\n<tr>\n<th style=\"text-align: left;\">Benchmark<\/th>\n<th style=\"text-align: left;\">Muse Glimmer-30B<\/th>\n<th style=\"text-align: left;\">Gemma4-31B<\/th>\n<th style=\"text-align: left;\">Qwen3.6-27B<\/th>\n<th style=\"text-align: left;\"><em>Kimi K3<\/em><\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td style=\"text-align: left;\">MBCT<\/td>\n<td style=\"text-align: left;\">41.5%<\/td>\n<td style=\"text-align: left;\"><strong>50.6%<\/strong><\/td>\n<td style=\"text-align: left;\">45.9%<\/td>\n<td style=\"text-align: left;\"><em>58.9%<\/em><\/td>\n<\/tr>\n<tr>\n<td style=\"text-align: left;\">HPCT<\/td>\n<td style=\"text-align: left;\">52.3%<\/td>\n<td style=\"text-align: left;\"><strong>54.0%<\/strong><\/td>\n<td style=\"text-align: left;\">48.7%<\/td>\n<td style=\"text-align: left;\"><em>59.6%<\/em><\/td>\n<\/tr>\n<tr>\n<td style=\"text-align: left;\">VCT<\/td>\n<td style=\"text-align: left;\">37.0%<\/td>\n<td style=\"text-align: left;\"><strong>43.5%<\/strong><\/td>\n<td style=\"text-align: left;\">33.7%<\/td>\n<td style=\"text-align: left;\"><em>48.0%<\/em><\/td>\n<\/tr>\n<tr>\n<td style=\"text-align: left;\">WMDP (Bio)<\/td>\n<td style=\"text-align: left;\"><strong>86.5%<\/strong><\/td>\n<td style=\"text-align: left;\">85.9%<\/td>\n<td style=\"text-align: left;\">84.8%<\/td>\n<td style=\"text-align: left;\"><em>89.1%<\/em><\/td>\n<\/tr>\n<tr>\n<td style=\"text-align: left;\">WMDP (Chem)<\/td>\n<td style=\"text-align: left;\">75.2%<\/td>\n<td style=\"text-align: left;\"><strong>80.5%<\/strong><\/td>\n<td style=\"text-align: left;\">74.8%<\/td>\n<td style=\"text-align: left;\"><em>84.2%<\/em><\/td>\n<\/tr>\n<tr>\n<td style=\"text-align: left;\">Lab Bench (ProtocolQA)<\/td>\n<td style=\"text-align: left;\"><strong>80.2%<\/strong><\/td>\n<td style=\"text-align: left;\">75.8%<\/td>\n<td style=\"text-align: left;\">69.1%<\/td>\n<td style=\"text-align: left;\"><em>81.9%<\/em><\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n<p>In the scientific field, it records 86.5% on WMDP (Bio) and 80.2% on Lab Bench (ProtocolQA), showing top-class performance among open-weight models of the same class. Meanwhile, in metrics such as MBCT, VCT, and WMDP (Chem), it lags behind Gemma4-31B, showing a tendency to have strengths and weaknesses depending on the field.<\/p>\n<h3>Verification of Performance Degradation Due to Quantization<\/h3>\n<p>The developers have verified the performance degradation when compressing the model (based on the average accuracy of 15 common benchmarks).<\/p>\n<div class=\"lmw-table-scroll\" tabindex=\"0\" style=\"overflow-x:auto;-webkit-overflow-scrolling:touch;max-width:100%;\">\n<table style=\"width:max-content;min-width:100%;border-collapse:collapse;\">\n<thead>\n<tr>\n<th style=\"text-align: left;\"><\/th>\n<th style=\"text-align: center;\">Full Precision<\/th>\n<th style=\"text-align: center;\">K-Quant-Dynamic<\/th>\n<th style=\"text-align: center;\">K-Quant-17GB<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td style=\"text-align: left;\">% Degradation*<\/td>\n<td style=\"text-align: center;\">&#8211;<\/td>\n<td style=\"text-align: center;\">0.2%<\/td>\n<td style=\"text-align: center;\">1.0%<\/td>\n<\/tr>\n<tr>\n<td style=\"text-align: left;\">Target Hardware<\/td>\n<td style=\"text-align: center;\">64GB VRAM<\/td>\n<td style=\"text-align: center;\">32GB VRAM<\/td>\n<td style=\"text-align: center;\">24GB VRAM<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n<p>According to this verification, even with the reduced model size &#8220;K-Quant-17GB&#8221;, the performance drop from full precision is kept to a mere 1.0%, and it is reported that practical accuracy can be maintained even on consumer-grade hardware such as a 24GB VRAM environment.<\/p>\n<h3>Inference Speed in Local Environments (DFlash Speculative Decoding)<\/h3>\n<p>Muse Glimmer bundles a lightweight speculative decoding model called &#8220;DFlash&#8221;, which significantly improves inference speeds in local environments. Below are the measurement results on various local hardware when combining the K-Quant-17GB model with the DFlash drafter:<\/p>\n<div class=\"lmw-table-scroll\" tabindex=\"0\" style=\"overflow-x:auto;-webkit-overflow-scrolling:touch;max-width:100%;\">\n<table style=\"width:max-content;min-width:100%;border-collapse:collapse;\">\n<thead>\n<tr>\n<th style=\"text-align: left;\">GPU<\/th>\n<th style=\"text-align: center;\">Baseline No-speculation (tok\/s)<\/th>\n<th style=\"text-align: center;\">Avg* with DFlash Speculation (tok\/s)<\/th>\n<th style=\"text-align: center;\">Speedup<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td style=\"text-align: left;\"><strong>Nvidia RTX 5090<\/strong><\/td>\n<td style=\"text-align: center;\">74.9<\/td>\n<td style=\"text-align: center;\">233.4<\/td>\n<td style=\"text-align: center;\">3.1x<\/td>\n<\/tr>\n<tr>\n<td style=\"text-align: left;\"><strong>Apple M4 Max<\/strong><\/td>\n<td style=\"text-align: center;\">23.7<\/td>\n<td style=\"text-align: center;\">37.8<\/td>\n<td style=\"text-align: center;\">1.5x<\/td>\n<\/tr>\n<tr>\n<td style=\"text-align: left;\"><strong>Apple M5 Max<\/strong><\/td>\n<td style=\"text-align: center;\">26.6<\/td>\n<td style=\"text-align: center;\">50.2<\/td>\n<td style=\"text-align: center;\">1.8x<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n<p>Compared to the baseline without speculative decoding, a 3.1x speedup (233.4 tok\/s) on the Nvidia RTX 5090 environment and a 1.8x speedup (50.2 tok\/s) on the Apple M5 Max environment are confirmed, making very smooth operation expected even on local devices.<\/p>\n<h2>Strengths and Use Cases<\/h2>\n<p>Muse-Glimmer-30B is designed with a primary focus on use as an autonomous AI agent running on consumer-grade hardware without relying on external networks or cloud infrastructure. According to the model card by Meta Superintelligence Lab, the developer of the original model, this model is optimized particularly for the following use cases and tasks:<\/p>\n<ul>\n<li><strong>Autonomous Agents in Local Environments<\/strong>: Equipped with multi-step long-term planning, sequential execution of ordered tools, and error self-diagnosis with retry capabilities upon runtime failures. It can be operated in a completely local environment without going through the cloud.<\/li>\n<li><strong>Software Development and Coding Agents<\/strong>: Supports SWE-Bench style workflows and excels at autonomously solving code creation, bug fixing, and real-world software engineering tasks.<\/li>\n<li><strong>Schema-Compliant Tool Use and Function Calling<\/strong>: Supports function calling based on strict schemas, allowing stable tool invocation even in complex workflows with multi-turn interactions.<\/li>\n<li><strong>Multimodal Reasoning<\/strong>: Integrates a dedicated perception encoder (ViT-G\/14) with approximately 1.8B parameters, enabling the interpretation of inputs combining text and images. Screenshots, charts, and document images can be read alongside conversational context.<\/li>\n<li><strong>Compatibility with Agent Frameworks<\/strong>: Supports major agent orchestration patterns such as OpenClaw and Hermes Agent.<\/li>\n<li><strong>Reasoning Strength Control<\/strong>: Reasoning depth can be adjusted through system prompts. For complex problem solving, coding, and agent tasks, specifying &#8220;high&#8221; or &#8220;xhigh&#8221; is recommended.<\/li>\n<li><strong>Multilingual Support<\/strong>: Trained on data in over 100 languages.<\/li>\n<li><strong>Synthetic Data Generation and LLM Evaluation<\/strong>: It is also envisioned for use in generating high-quality training data for subsequent model development or as an LLM-as-a-judge to evaluate outputs from other models.<\/li>\n<\/ul>\n<p>Note that audio input and output are not supported, and video input is processed as individual still image frames. Furthermore, downloading or use by users under 18 years of age is not intended.<\/p>\n<p>The sampling settings recommended by the developers are as follows:<\/p>\n<ul>\n<li><strong>temperature<\/strong>: 1.0<\/li>\n<li><strong>top_p<\/strong>: 0.95<\/li>\n<li><strong>top_k<\/strong>: 64<\/li>\n<li><strong>Reasoning strength<\/strong>: <code>Reasoning strength: high<\/code> or <code>xhigh<\/code> (specified as low \/ medium \/ high \/ xhigh within the system prompt)<\/li>\n<\/ul>\n<p><!-- lmw:hardware --><\/p>\n<h2>Hardware Requirements<\/h2>\n<p><strong>Estimated requirements (calculated by Local Model Watch)<\/strong> \u2014 29.8B parameters (taken from the base model meta-models\/Muse-Glimmer-30B)<\/p>\n<div class=\"lmw-table-scroll\" tabindex=\"0\" style=\"overflow-x:auto;-webkit-overflow-scrolling:touch;max-width:100%;\">\n<table style=\"width:max-content;min-width:100%;border-collapse:collapse;\">\n<thead>\n<tr>\n<th>Your VRAM<\/th>\n<th>Quantization<\/th>\n<th>File size<\/th>\n<th>Est. memory needed<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>24GB (RTX 4090 \/ 3090, etc.)<\/td>\n<td>Q4_K_M<\/td>\n<td>17.9GB<\/td>\n<td>21.5GB<\/td>\n<\/tr>\n<tr>\n<td>48GB (RTX 6000 Ada \/ A6000, etc.)<\/td>\n<td>Q8_0<\/td>\n<td>27.6GB<\/td>\n<td>33.1GB<\/td>\n<\/tr>\n<tr>\n<td>80GB class (A100 \/ H100)<\/td>\n<td>BF16<\/td>\n<td>51.9GB<\/td>\n<td>62.3GB<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n<p><em>Memory estimates add a 20% runtime overhead (KV cache, etc.) to the actual size of the distributed files. Actual usage varies with context length, batch size and inference engine. These figures are computed by this site from file sizes, not published by the model&#8217;s authors. Compare with other models in our <a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/vram-guide-en\/\">VRAM quick reference<\/a>. What the quantization names mean: <a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/glossary-quantization-en\/\">glossary<\/a>.<\/em><\/p>\n<p><!-- \/lmw:hardware --><\/p>\n<p><!-- lmw:runnability --><\/p>\n<h2>Can You Run It Locally?<\/h2>\n<p><strong>Runs in Ollama, LM Studio and llama.cpp as-is.<\/strong><\/p>\n<p>It is distributed in GGUF, so no conversion is needed.<\/p>\n<p><strong>License \u2014 <code>apache-2.0<\/code> (Commercial use allowed):<\/strong> Permits commercial use, modification and redistribution. Redistribution requires including the license and stating changes; includes a patent grant.<\/p>\n<p><strong>Compression:<\/strong> the Q4_K_M build measures 5.16 bits per weight \u2014 about 32% the size of the original 16-bit weights, calculated by this site from the actual file sizes.<\/p>\n<p><em>Compiled by this site&#8217;s code from the published formats, converted builds we have found, and each engine&#8217;s own model registry. &#8220;Not found&#8221; means we have not seen such a build, not that none exists. License summaries are not legal advice \u2014 check the publisher&#8217;s original terms before relying on them.<\/em><\/p>\n<p><!-- \/lmw:runnability --><\/p>\n<p><!-- lmw:peers --><\/p>\n<h2>Recent Models in the Same Size Class<\/h2>\n<p><em>Models with <\/em><em>15\u201340B<\/em><em> parameters that Local Model Watch covered recently, listed by code from our article log for comparison. VRAM tiers are this site&#8217;s estimates; licenses are as stated on the model cards.<\/em><\/p>\n<div class=\"lmw-table-scroll\" tabindex=\"0\" style=\"overflow-x:auto;-webkit-overflow-scrolling:touch;max-width:100%;\">\n<p class=\"lmw-table-hint\" style=\"margin:0 0 4px;font-size:0.85em;opacity:0.7;\">\u2192 Scroll horizontally to see all columns<\/p>\n<table style=\"width:max-content;min-width:100%;border-collapse:collapse;\">\n<thead>\n<tr>\n<th>Model<\/th>\n<th>Parameters<\/th>\n<th>Smallest VRAM tier<\/th>\n<th>License<\/th>\n<th>Our article<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>Cloudflare\/clef-omni<\/td>\n<td>35.3B<\/td>\n<td>80GB<\/td>\n<td>apache-2.0<\/td>\n<td><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/11\/17\/cloudflare-releases-clef-omni\/\">clef-omni 30B Multimodal Decision Model: 80GB+ VRAM<\/a> (2026-11-17)<\/td>\n<\/tr>\n<tr>\n<td>ggml-org\/vllm-d3-GGUF<\/td>\n<td>26.1B<\/td>\n<td>24GB<\/td>\n<td>apache-2.0<\/td>\n<td><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/11\/17\/vllm-d3-gguf-models-released\/\">vllm-d3 GGUF Models Released: Decision Models for Local Run<\/a> (2026-11-17)<\/td>\n<\/tr>\n<tr>\n<td>inclusionAI\/ZwZ2-27B-GGUF<\/td>\n<td>27.4B<\/td>\n<td>24GB<\/td>\n<td>\u2014<\/td>\n<td><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/11\/17\/zwz2-27b-gguf-released\/\">ZwZ2-27B-GGUF Vision-Language Model: 24GB+ VRAM<\/a> (2026-11-17)<\/td>\n<\/tr>\n<tr>\n<td>Cloudflare\/clef<\/td>\n<td>27.4B<\/td>\n<td>12GB<\/td>\n<td>apache-2.0<\/td>\n<td><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/10\/02\/cloudflare-clef-27b-multimodal-decision-model\/\">clef Structured Decision-Making Model: 12GB+ VRAM, GGUF Builds<\/a> (2026-10-01)<\/td>\n<\/tr>\n<tr>\n<td>ggml-org\/OpenJev-GGUF<\/td>\n<td>27.4B<\/td>\n<td>24GB<\/td>\n<td>cc-by-nc-4.0<\/td>\n<td><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/10\/02\/ggml-org-openjev-gguf-released-2\/\">OpenJev-GGUF 27.4B Decision-Making Specialized Model: 24GB+ VRAM<\/a> (2026-10-01)<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n<p><!-- \/lmw:peers --><\/p>\n<h2>How to Get It<\/h2>\n<p>&#8220;ggml-org\/Muse-Glimmer-30B-GGUF&#8221; is available on Hugging Face, and since it is not a gated model, you can download it directly without prior application or consent procedures. The Apache 2.0 license applies, just like the original model.<\/p>\n<h3>Distribution Format<\/h3>\n<p>Model files are provided in GGUF format in this repository. In addition to the main model, the following components are provided:<\/p>\n<ul>\n<li><strong>Projector for Vision Encoder<\/strong>: mmproj file with <code>Q8_0<\/code> quantization<\/li>\n<li><strong>Speculative Decoding Drafter (DFlash drafter sidecar)<\/strong>: <code>BF16<\/code>, <code>Q8_0<\/code>, and <code>Q4_0<\/code> variants<\/li>\n<\/ul>\n<p>This model conversion is automatically performed using ggml-org&#8217;s conversion tool (<a href=\"https:\/\/github.com\/ggml-org\/convert\">https:\/\/github.com\/ggml-org\/convert<\/a>).<\/p>\n<h3>How to Run<\/h3>\n<p>It is compatible with the llama.cpp ecosystem, allowing you to easily start a server from environments such as llama.app. The official model card provides instructions on how to run it using the following command:<\/p>\n<pre><code class=\"language-bash\">llama serve -hf ggml-org\/Muse-Glimmer-30B-GGUF\n<\/code><\/pre>\n<p>By combining the DFlash drafter model to enable speculative decoding, generation speeds can be significantly boosted compared to single-token generation.<\/p>\n<p><!-- lmw:related --><\/p>\n<h2>Related Articles<\/h2>\n<ul>\n<li><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/11\/17\/qwen3-coder-30b-a3b-instruct-gguf-2\/\">Qwen3-Coder-30B-A3B-Instruct-GGUF Text Generation Model: 24GB+ VRAM<\/a><\/li>\n<li><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/11\/17\/vllm-d3-gguf-models-released\/\">vllm-d3 GGUF Models Released: Decision Models for Local Run<\/a><\/li>\n<li><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/10\/02\/ggml-org-openjev-gguf-released-2\/\">OpenJev-GGUF 27.4B Decision-Making Specialized Model: 24GB+ VRAM<\/a><\/li>\n<li><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/11\/17\/qwen3-5-4b-gguf-released\/\">Qwen3.5-4B-GGUF Vision-Language Model: 4GB+ VRAM<\/a><\/li>\n<\/ul>\n<p><!-- \/lmw:related --><\/p>\n<p><!-- lmw:next-steps --><\/p>\n<h2>What to Read Next<\/h2>\n<ul>\n<li><strong>Find models by VRAM<\/strong> (This model runs from the 24GB tier) \u2192 <a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/vram-guide-24gb-en\/\">Other models that run on a 24GB GPU<\/a><\/li>\n<li><strong>Engines that run this model<\/strong> \u2192 <a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/engine-llama-cpp-en\/\">llama.cpp<\/a> \/ <a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/engine-ollama-en\/\">Ollama<\/a><\/li>\n<li><strong>What Q4_K_M, Q8_0 mean and where to get this model<\/strong> \u2192 <a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/format-gguf-en\/\">GGUF format guide and models<\/a> \/ <a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/glossary-quantization-en\/\">Quantization and model-format glossary<\/a><\/li>\n<li><strong>Other models for the same task<\/strong> \u2192 <a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/models-by-task-en\/#task-vision\">Other vision-language models<\/a><\/li>\n<li><strong>How to read OSWorld, SWE-bench, Terminal-Bench<\/strong> \u2192 <a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/glossary-benchmarks-en\/\">Benchmark glossary<\/a><\/li>\n<\/ul>\n<p><!-- \/lmw:next-steps --><\/p>\n<h2>Sources<\/h2>\n<ul>\n<li><a href=\"https:\/\/huggingface.co\/ggml-org\/Muse-Glimmer-30B-GGUF\">https:\/\/huggingface.co\/ggml-org\/Muse-Glimmer-30B-GGUF<\/a><\/li>\n<li><a href=\"https:\/\/huggingface.co\/meta-models\/Muse-Glimmer-30B\">https:\/\/huggingface.co\/meta-models\/Muse-Glimmer-30B<\/a><\/li>\n<\/ul>\n","protected":false},"excerpt":{"rendered":"<p>GGUF quantization of Meta&#8217;s Muse-Glimmer-30B multimodal local agent model is now available, featuring local execution and tool use.<\/p>\n","protected":false},"author":1,"featured_media":10706,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[310],"tags":[161,1970,163,3166,165,3168,3170,1547,1952],"class_list":["post-10707","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-new-models","tag-ggml-en","tag-ggml-org-en","tag-gguf-en","tag-meta-en","tag-moe-en","tag-muse-glimmer-en","tag-muse-glimmer-30b-en","tag-verified","tag-vlm-en"],"lang":"en","translations":{"en":10707,"ja":10705},"_links":{"self":[{"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts\/10707","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/comments?post=10707"}],"version-history":[{"count":2,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts\/10707\/revisions"}],"predecessor-version":[{"id":17723,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts\/10707\/revisions\/17723"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/media\/10706"}],"wp:attachment":[{"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/media?parent=10707"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/categories?post=10707"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/tags?post=10707"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}