muse-glimmer — imatrix quantizations for local hardware

GGUF builds of Meta’s Muse Glimmer 30B, re-quantized down to 2-bit using an importance-matrix (imatrix) calibration pass so the smaller builds keep as much reasoning and tool-use quality as possible for their size. base is Meta’s official K-Quant-17GB release, unmodified, kept here as the reference point for the compressed builds.

About Muse Glimmer

Muse Glimmer is a 30B-parameter (29.6B incl. its vision encoder) open-weight model released by Meta Superintelligence Labs in August 2026 under the Apache 2.0 license — distilled from Muse Spark and purpose-built for agentic workloads: multi-step reasoning, reliable schema-based tool calling, failure recovery, and native multimodal (text + image) input through a dedicated 1.8B-parameter perception encoder. It supports a 128K context window and four controllable reasoning-effort levels (low, medium, high, xhigh) so you can trade speed for depth per task.

Original weights: meta-models/Muse-Glimmer-30B GGUF source: meta-models/Muse-Glimmer-30B-GGUF

What’s in this repo

All non-base builds were requantized from the full-precision weights using an importance matrix computed over a diverse calibration set (general text, code, multi-turn tool-call traces) so that the weights that matter most for reasoning and function-calling are preserved at higher effective precision, even though the average bits-per-weight is ~2. Context length is preserved at 128K across every build.

Quick start

ollama run venkataparswanadh/muse-glimmer:base      # best quality, needs ~24GB

ollama run venkataparswanadh/muse-glimmer:iq2_m     # best small-model pick, ~10GB

ollama run venkataparswanadh/muse-glimmer:q2_k      # ~11GB

ollama run venkataparswanadh/muse-glimmer:iq2_xs    # smallest, ~9GB

Choosing a build

• Have 24GB+ VRAM or unified memory? Use base. It’s Meta’s own quant, unmodified — the compressed builds exist to make Muse Glimmer runnable below that line, not to replace it.
• 12–16GB? Start with iq2_m. In testing it held up noticeably better than a naive 2-bit quant on multi-step tool-calling and long-context coherence, at a smaller footprint than q2_k.
• 8–10GB, or want the fastest tokens/sec? iq2_xs is the floor — expect some quality loss on harder agentic/coding tasks, less on everyday chat.

At 2-bit, arithmetic and precise multi-step tool schemas degrade first; general reasoning and conversation hold up better. For anything tool-call-heavy or code-heavy, prefer base or iq2_m if you have the memory.

Reasoning effort

Muse Glimmer supports controllable reasoning strength. Set it in your prompt or system message as low / medium / high / xhigh — use high/xhigh for coding and agentic tasks, low/medium when you just want fast answers.

Recommended sampling

Temperature 1.0, top-p 0.95, top-k 64 (Meta’s published defaults).

License & attribution

Apache 2.0, inherited from the original Meta Superintelligence Labs release. This repo redistributes Meta’s official K-Quant build (base) unmodified, and ships derivative quantizations of Meta’s weights (q2_k, iq2_xs, iq2_m) produced independently by me — not affiliated with or endorsed by Meta. See Meta’s Usage Policy for acceptable-use terms that still apply to these derivatives.

Contact & feedback

Found a bug, have a hardware setup that struggles with one of the builds, or just want to say it worked well for you? Reach out:

📧 venkataparswanadh@gmail.com

Reviews, benchmarks from your own hardware, and general feedback are all genuinely appreciated — they help decide which quant levels are worth maintaining going forward.