Muse-Glimmer-30B: unknown model architecture 'muse-glimmer' in b10269-1.5.1 (upstream PR #26841 has the port)

[itsmrshow]

Summary

Meta released Muse-Glimmer-30B (2026-08-09) — a dense 30B multimodal reasoning model with a DFlash block-diffusion drafter. The current release b10269-1.5.1 (2026-08-06) predates it, so loading the official GGUF fails with:

unknown model architecture: 'muse-glimmer'

Upstream status

The arch landed in upstream llama.cpp via PR ggml-org#26841 "model: Muse Glimmer Support" (merged 2026-08-10, merge commit 62bf73d2, 22 files / +877). It includes:

• LLM_ARCH_MUSE_GLIMMER (dense 30B, 52 layers, [Local×3, Global] sliding-window pattern, GQA 16:1, RoPE theta 500K local-only)
• Full DFlash drafter support (dflash arch + --spec-type draft-dflash), which pairs naturally with the fork's existing DFlash implementation
• Multimodal mmproj support (ViT-G/14 perception encoder) in tools/mtmd/
• Chat template handler for the <|start|>assistant to=<recipient><|message|> format with ATEM tool-call markup

Note for the merge

Since the fork tracks upstream via merge commits, a normal upstream catch-up merge should bring this in — no cherry-pick needed. The DFlash pieces in the fork (LLM_ARCH_DFLASH, common_speculative_impl_draft_dflash) are already present and compatible; upstream's speculative.cpp change in the PR is a small embedding-batch fix that enables DFlash on multimodal targets.

Reference

• Model card: meta-models/Muse-Glimmer-30B (GGUF: meta-models/Muse-Glimmer-30B-GGUF)
• Upstream PR: model: Muse Glimmer Support ggml-org/llama.cpp#26841
• I've also verified a local port of PR model: Muse Glimmer Support ggml-org/llama.cpp#26841 onto b10269-1.5.1 works (model loads, DFlash + mmproj + turbo4 KV all functional) — happy to share the patch if useful.