hipfire-models/muse-glimmer-30b

Muse Glimmer 30B — hipfire quants

MQ4 quantizations of meta-models/Muse-Glimmer-30B for hipfire, an inference engine for AMD RDNA / CDNA GPUs.

These files require hipfire. They are not GGUF and will not load in llama.cpp, transformers, vLLM, or Ollama.

Text-only. Upstream Muse Glimmer is a vision-language model (image-text-to-text). These artifacts contain the dense text tower only — 418 weight tensors (52 layers × 5 attention + 52 × 3 MLP + embed + lm_head). There is no perception encoder in these files and image input is not supported.

Files

file	size	attention	lm_head	notes
muse-glimmer-30b.mq4	18.6 GB	Q8	Q8	quality trunk
muse-glimmer-30b.mq4r	16.3 GB	MQ4	Q8	speed SKU, 2.35 GB smaller
muse-glimmer-30b-dflash.mq4	1.36 GB	—	—	DFlash draft, pairs with either SKU

All three share an MQ4 body; they differ in attention precision.

sha256  87bc776a36e0c4b55edfd1f0da1104c966eb43f59aa6fdd0ee33a14ecc0fc052  muse-glimmer-30b.mq4
sha256  47ccfccddbef5b8e14040bae567c87129669fa9451f0c6b15d8f35964fce7503  muse-glimmer-30b.mq4r
sha256  6fc0988e51689abf24c2e43d5b1f44794be3409484f99a115d36484ea0e23d2d  muse-glimmer-30b-dflash.mq4

Usage

hipfire pull muse-glimmer          # quality trunk (.mq4)
hipfire pull muse-glimmer:fast     # speed SKU (.mq4r)
hipfire pull muse-glimmer:draft    # DFlash draft

hipfire run muse-glimmer "Write a merge sort in Python"

Speculative decode needs an explicit draft path

Unlike hipfire's Qwen models, Glimmer has no filename auto-pairing. Pulling the draft is not enough — the daemon will run plain autoregressive decode and you will conclude DFlash is broken. Attach it explicitly:

export HIPFIRE_DFLASH_DRAFT=~/.hipfire/models/muse-glimmer-30b-dflash.mq4
hipfire config set dflash_mode auto

The draft is a 5-layer block-diffusion head (muse_glimmer_assistant, arch 23) that reuses the target's embed and lm_head.

Which SKU

.mq4 is the default. Take .mq4r if you are tight on VRAM — it saves 2.35 GB and decodes a little faster, at MQ4 attention precision.

Measured on gfx1201 (Radeon AI PRO R9700), 2026-08-14:

SKU	AR decode	with DFlash
.mq4	28.4 tok/s	225 tok/s
.mq4r	32.3 tok/s	229 tok/s

The AR gap is ~14%; under speculative decode the two nearly converge. Your numbers will differ — prompt structure alone can move hipfire throughput by double digits, so treat these as a shape, not a guarantee.

.mq4r is not Redline-lowered

The r suffix marks an MQ4R Redline SKU, but Glimmer is not lowered to Redline PM4 yet, so automatic Redline admission is deliberately withheld. The suffix describes the format, not an active fast path.

Quality

Not measured. These are speed/size SKUs published without a KLD or perplexity number for this architecture. hipfire has no arch-14 evaluator and no Glimmer reference, so no quality claim is made here — in either direction.

Both SKUs were validated for coherence (a five-genre battery, plus speculative-decode agreement against the trunk), not for divergence from full precision. If you need a measured quant, wait for one.

Requirements

• AMD GPU: gfx11 (RDNA3) or gfx12 (RDNA4) recommended; gfx1201 is the reference
• ~26 GB VRAM for .mq4, ~24 GB for .mq4r, at the default q8 KV mode
• ROCm, and a hipfire build with arch-muse-glimmer

License and attribution

Apache-2.0, inherited from upstream Muse Glimmer. These are quantized derivatives — the weights are Meta Models' work; the quantization formats (MQ4 / FWHT-rotated Magnum Quant), the DFlash draft, and the engine are hipfire's. See NOTICE and PRIOR-ART.md.