Model datasheet · Muse
Muse Glimmer 30B
Meta's agent-first comeback under a real Apache license: strong tool use, screen understanding, and a speculative-decoding sidecar for speed. Knowledge is its weak leg — it's built to act, not to recall.
- Vendor
- Meta
- Architecture
- Dense · 29.6B
- Context
- 131,072 tokens
- License
- Apache-2.0
- Released
- 2026-08-10
- Vision
- Yes
→ GGUF (llama.cpp / LM Studio / Ollama) → MLX (Apple Silicon)
Architecture, from its own config.json
52 layers · 39 sliding window · 13 full attention
32 attention heads over 2 KV heads (GQA 16:1), head dim 128 · dense feed-forward · vocabulary 202,048
Where the memory goes: weights by band above, plus a KV cache of 202MB at 8K · 514MB at 32K · 1.7GB at 128K, plus the runtime's own working memory — dominated by the output projection over that vocabulary.
Inside one layer, from its own tensors
52 × layer
39 sliding-window layers, 13 full-attention layers — the same block either way. What differs is how far back the mask lets attention look, not what the layer is built from.
The green line down the right of each half is the residual: it leaves the spine above the norm, skips the operation entirely and rejoins at the +. Where attention is drawn open, q, k and v are scaled to their real output widths — a model that shares key and value heads across many query heads shows one wide box beside two narrow ones.
Around the stack: an embedding of 202,048 × 6,656, a final norm, and a separate output projection.
§1 Characteristics by quantization band
— Not yet rated · contributions welcome —
Q4–Q5 — 4–5 bit
17GB weights · e.g. Q4_K_M , mlx-4bit + KV: 202MB at 8K · 514MB at 32K · 1.7GB at 128Kdata checked Aug 2026
- quant-degradation-community — KLD/perplexity delta vs bf16 @ Q4_K_M · aggregated · 2026-08-30
Editorial derivation: small loss at 4-bit for a 30B dense model.
data checked Aug 2026
- quant-degradation-community — KLD/perplexity delta vs bf16 @ Q4_K_M · aggregated · 2026-08-30
Editorial derivation: small loss at 4-bit.
data checked Aug 2026
- quant-degradation-community — community consensus @ Q4_K_M · aggregated · 2026-08-30
Editorial derivation: tool-call formatting holds at 4-bit; official Q4_K_M is Meta's own recommended local build.
data checked Aug 2026
- quant-degradation-community — community consensus @ Q4_K_M · aggregated · 2026-08-30
Editorial derivation: IF is not quant-sensitive.
data checked Oct 2026
- quant-degradation-community — editorial band derivation @ Q4_K_M · aggregated · 2026-10-04
Editorial derivation: carried from the EQ-Bench-anchored full band; creative writing is not quant-sensitive.
— Not yet rated · contributions welcome —
FP16 — full precision (fp16/bf16)
59GB weights · e.g. bf16 + KV: 202MB at 8K · 514MB at 32K · 1.7GB at 128Kdata checked Aug 2026
- vendor-model-card — AIME 2026 = 94.7 @ bf16 · vendor-reported · 2026-08-30
GPQA-Diamond: 83.5. No independent math leaderboard lists it: absent from LiveBench (release 2026-06-25) as of 2026-10-04. Aider polyglot has not added an entry since 2025-10-03, so its absence there says nothing.
data checked Aug 2026
- vendor-model-card — SWE-bench Verified = 76 @ bf16 · vendor-reported · 2026-08-30
SWE-bench Pro: 51.2. Community judges its agentic coding stronger than its raw single-file coding.
data checked Aug 2026
- vendor-model-card — MCP Atlas = 75.5 @ bf16 · vendor-reported · 2026-08-30
TerminalBench 2.1: 51.7, ScreenSpot Pro: 75.4 — the agentic story is why this model exists.
data checked Aug 2026
- vendor-model-card — IFBench = 77 @ bf16 · vendor-reported · 2026-08-30
data checked Aug 2026
- artificial-analysis — AA-Omniscience (value at source — their terms bar republication) @ bf16 · aggregated · 2026-08-30
Secondary coverage of AA's knowledge eval reports very high hallucination on hard recall; see the known issue below for citable analysis.
data checked Oct 2026
- eqbench-creative — EQ-Bench Creative Writing v3 rubric score (0-20) = 16.26 @ bf16 · aggregated · 2026-10-04
Independent; rubric judged by Claude Sonnet 4. Normalised Elo 1798.3, 26th of 140 and the highest-placed model on this site, retrieved 2026-10-04 (Elo moves as models are added, so the rubric score is the figure stored). Ahead of Qwen3.8 27B (15.50), every Gemma 4 model (15.71-16.03) and Gemma 3 27B (15.33). Not on the Longform board. Quant not stated by EQ-Bench (presumed full-precision API serving). Data file behind https://eqbench.com/creative_writing.html.
§2 Known issues & what fixes them
Hallucinates heavily on hard knowledge questions — independent knowledge evaluations report heavy hallucination on hard recall, a stark contrast with its agentic strength.
It was built for tool use — give it search and it stops guessing.
Grounded retrieval covers document work.
Evidence · 2 sources · community-consensus
- Muse Glimmer 30B benchmark and hardware analysis (blog, 2026-08-30)
- BenchLM Muse Glimmer 30B profile (blog, 2026-08-30)
A GGUF converted from the NVFP4 (compressed-tensors) checkpoint fails to load, and once patched past the load error returns the same token for every prompt. Reproduced on CUDA and on Apple silicon; still open as of 2026-10-04. K-quant conversions of the same model work, so this only bites if you pick an NVFP4 file.
Use a K-quant GGUF instead, such as Meta's own Q4_K_M: same band, and unaffected.
Evidence · 1 source · anecdotal
- Eval bug: Muse Glimmer NVFP4 fails to load, then emits one repeated token (open) (github, 2026-10-04)