Muse Glimmer 30B - pruned V2.5 EN-PL GGUF

Pruned GGUF build of meta-models/Muse-Glimmer-30B, published by bluevoid-pl.

• Base model: meta-models/Muse-Glimmer-30B
• Size: 27B params
• Architecture: muse-glimmer
• License: Apache 2.0
• HF: bluevoid-pl/Muse-Glimmer-30B-pruned-V2.5-en-pl-GGUF

Overview

Muse Glimmer is a 30B causal language model with a dedicated perception encoder, distilled from Muse Spark and purpose-built for autonomous agentic tasks on consumer hardware. The model integrates multi-step reasoning, reliable tool use, multimodal understanding, and failure recovery into a single model that runs locally without requiring cloud infrastructure or network access.

This V2.5 build adds vision support and prunes to English + Polish characters only.

Pruning

This GGUF is pruned:

• Only English and Polish characters are kept, so emojis and non-latin scripts are physically removed (model still thinks it can output them).
• Pruning reduces model size by ~25%, so it fits on limited VRAM.
• No guarantee of any kind that this GGUF works at all. It might break or die for no reason.

Vision capabilities are restored compared to previous versions. Vision requires downloading mmproj-* file. Quantization of vision models usually compound error, so it's recommended to use Q8_0 for vision regardless of your text-text quantization level. Correct file is: mmproj-Muse-Glimmer-30B_pruned-pl-V2.5-Q8_0.gguf.

Note: This model is using English thinking, because of that using Polish language in your prompt is not ideal.

Running

Requires llama.cpp newer than b10430; older versions fail to load. Support in llama.cpp and ollama is partial - prefer the web interface over the CLI. On first start llama.cpp may hang for ~5min and warn special_eot_id is not in special_eog_ids - that warning is expected.

Text-only (no vision)

llama.cpp CLI

./llama-cli -m Muse-Glimmer-30B-pruned-V2.5-en-pl-IQ3_XXS.gguf -n 512 --temp 0.6 --jinja

llama-server for 12Gb VRAM

./llama-server -m Muse-Glimmer-30B-pruned-V2.5-en-pl-IQ3_XXS.gguf --alias "default" --ctx-size 100000 --threads 6 --jinja -ngl 53 --host 0.0.0.0 --port 13378 --api-key "APIKEY" --parallel 1 --temp 0.6 --reasoning-preserve

Ollama

ollama run hf.co/bluevoid-pl/Muse-Glimmer-30B-pruned-V2.5-en-pl-GGUF:IQ3_XXS

With vision

Vision requires the mmproj file: mmproj-Muse-Glimmer-30B_pruned-pl-V2.5-Q8_0.gguf. Quantization of vision models compounds error, so use Q8_0 for vision regardless of text quantization.

llama.cpp CLI

./llama-cli -m Muse-Glimmer-30B-pruned-V2.5-en-pl-Q8_0.gguf --mmproj mmproj-Muse-Glimmer-30B_pruned-pl-V2.5-Q8_0.gguf -n 512 --temp 0.6 --jinja

llama-server

./llama-server -m Muse-Glimmer-30B-pruned-V2.5-en-pl-Q8_0.gguf --mmproj mmproj-Muse-Glimmer-30B_pruned-pl-V2.5-Q8_0.gguf --alias "default" --ctx-size 100000 --threads 6 --jinja -ngl 53 --host 0.0.0.0 --port 13378 --api-key "APIKEY" --parallel 1 --temp 0.6 --reasoning-preserve

Ollama (vision support is experimental)

ollama run hf.co/bluevoid-pl/Muse-Glimmer-30B-pruned-V2.5-en-pl-GGUF:Q8_0

Recommendations for vision

• Use Q8_0 quantization for the model and mmproj to minimize vision errors.
• Keep mmproj file alongside the GGUF and ensure llama.cpp >= b10430.
• For high-resolution images, increase --ctx-size and reduce --parallel to avoid OOM.
• Vision performance is best with --jinja and --reasoning-preserve enabled.

Quantizations