16 1 week ago

Meta's Apache 2.0 agentic vision-language model, 30B at 4-bit with 131K context

vision 30b
ollama run odytrice/muse

Models

View all →

Readme

Muse

Muse Glimmer profile for Ollama, published as odytrice/muse.

The card name is shortened to muse; the upstream model is Meta’s Muse Glimmer 30B.

This card does not use the <gpu>-<size> tag scheme the other cards use. A per-GPU split would be meaningless here — context is capped by the source GGUF rather than by VRAM, so a 24 GB and a 32 GB card get the identical profile.

Meta’s Apache 2.0 agentic model — distilled from Muse Spark and built for local tool use, multi-step reasoning, and failure recovery rather than for general chat.

Tags

Tag Quantization num_ctx
odytrice/muse:30b Ollama 4-bit (~18 GB) 131072
odytrice/muse:latest Ollama 4-bit (~18 GB) 131072

Both tags are the same build from 30b/Modelfile; latest is the convenience alias so ollama run odytrice/muse works.

Ollama publishes only 30b (~18 GB) and 30b-mlx (~21 GB, Apple silicon); there is no q8_0 or bf16 tag to build a higher-precision variant from the way gemma4:5090-12b has one.

Upstream

Size Upstream Architecture Modalities Native context
30B meta-models/Muse-Glimmer-30B Dense causal transformer, 52 layers, hidden 6656, GQA 32 query / 2 KV heads, head dim 128 Text + Image in, text out (~1.8B ViT-G/14 encoder) 131072 (documented up to 262144)

~29.6B parameters, Apache 2.0.

Environment

Always set these before running Ollama:

set OLLAMA_KV_CACHE_TYPE=q8_0
set OLLAMA_FLASH_ATTENTION=1

export OLLAMA_KV_CACHE_TYPE=q8_0
export OLLAMA_FLASH_ATTENTION=1

Context Size

The 32:2 GQA ratio makes the KV cache unusually cheap: 2 KV heads x 128 head dim x 2 (K and V) x 52 layers is ~26 KB per token at q8_0 cache, so 131072 tokens costs roughly 3.5 GB and 262144 costs roughly 7 GB. Against ~18 GB of weights that leaves the 4090 comfortable at full native context — unlike the 27B dense profiles, this model is not context-constrained on 24 GB.

131072 is a hard ceiling rather than a tuning choice. The packaged GGUF reports muse-glimmer.context_length: 131072, and Ollama silently clamps any larger num_ctx to it — a profile built with 262144 still loaded at 131072 under ollama ps. Meta documents the architecture as supporting up to 262144, but that requires RoPE scaling this GGUF does not carry (rope.freq_base 500000).

This is why there is one profile instead of a per-GPU pair: the extra VRAM on a 32 GB card buys nothing here. The genuine upgrade for a larger card would be a higher-precision build — Unsloth’s UD-Q6_K_XL at 20-22 GB — which Ollama’s library does not publish; see Notes.

Verified on a 5090: 16 GB resident, 100% GPU, no CPU offload at 131072. If ollama ps does show CPU offload, drop to 98304.

Sampling

Meta’s published defaults:

temperature 1.0
top_p       0.95
top_k       64

Note top_k 64 — higher than the top_k 20 used by the Qwen 3.6 / 3.8 profiles. Carrying a Qwen client config across will be wrong.

The model exposes adjustable reasoning effort (low, medium, high, xhigh) to trade compute against answer depth. These profiles do not pin a level; set it per request.

Notes

The profiles build on the default muse-glimmer:30b Ollama tag rather than a direct Hugging Face GGUF import, following the same compatibility precedent as the Qwen 3.6 / 3.8 profiles.

Unsloth publishes a UD-Q6_K_XL build at 20-22 GB that would fit a 5090 with room for KV cache and would be a genuine quality upgrade over 4-bit. It is a direct HF GGUF import, which this repo has avoided after malformed template and projector artifacts on earlier Ollama releases — worth testing, but not what these profiles ship.