GGUF builds of meta-models/Muse-Glimmer-30B — Meta Superintelligence Lab's ~30B multimodal agent model, distilled from Muse Spark and purpose-built for autonomous agentic work on consumer hardware. Every file here was cut from the same BF16 source weights; the repo holds full-precision copies, a vision projector, and 29 quants from Q8_0 down to Q1_0. Quantized 2026-09-07.

These are llama.cpp GGUF files. They are not loaded by vLLM, SGLang or transformers the way the source repo is.

Which file should I grab?

Default for most people: Q4_K_M (15.77 GB) — best quality-per-byte, in the same size class Meta targeted for mainstream local deployments.
Quality first: Q8_0 (27.58 GB) is effectively lossless; Q6_K (21.3 GB) is near-lossless.
Tighter on memory: IQ4_XS (14.29 GB) usually beats Q4_K_S at a smaller size; IQ3_M (11.94 GB) is the strong ~3.7 bpw pick; IQ2_M (9.17 GB) is the smallest size most people still find usable.
Legacy Q4_0 / Q4_1 / Q5_0 / Q5_1 / Q2_0 / Q1_0 are included for completeness or older runtimes — prefer the K-quants otherwise.
IQ2_XS, IQ2_XXS, IQ1_S, IQ1_M are extreme and experimental; expect substantial degradation.

About Muse Glimmer

Muse Glimmer is a 30B-class dense causal transformer with a dedicated perception encoder (image + text input, text output), trained for the capabilities an autonomous agent needs working together:

End-to-end agentic task completion — works inside scaffolds, writes and debugs code, resolves multi-turn requests from start to finish.
Reliable tool use — precise schema-based function calls across extended workflows.
Multi-step reasoning and failure recovery — sustains long-horizon plans; when a tool call fails, diagnoses and retries instead of halting.
Multimodal input — interleaved text and images: screenshots, charts, documents; up to 4,096 visual tokens per image.
Controllable effort — reasoning strength set in the system prompt (low / medium / high / xhigh).
Multilingual — trained on data from more than 100 languages.

Benchmarks reported by Meta (High Reasoning; compared against Gemma4-31B and Qwen3.6-27B Thinking Mode):

Benchmark	Muse Glimmer-30B
MCP Atlas (Public)	75.5
DeepSearch QA	74.6
SWE-Bench Pro	51.2
SWE-Bench Verified	76.0
AIME 2026	94.7
GPQA Diamond (AA)	83.5

Full tables and methodology are on the source model card and Meta's evaluation report.

Note on speed: Meta ships a DFlash speculative-decoding drafter with the source release (up to 3.1× faster generation on an RTX 5090 in their measurements). That drafter is not part of this GGUF repo — it contains only the main model and its vision projector.

Files

Full precision and the vision projector:

File	Size	Notes
Muse-Glimmer-30B-BF16.gguf	51.9 GB	Full-precision source; every quant below is cut from this file.
Muse-Glimmer-30B-F16.gguf	51.9 GB	Full-precision source.
mmproj-Muse-Glimmer-30B-F16.gguf	3.58 GB	Vision projector (F16). Download this too if you want image input — pass it to `llama-cli` / `llama-server` as the projector (`--mmproj`).

Quantized (largest → smallest):

File	Size	Notes
Muse-Glimmer-30B-Q8_0.gguf	27.58 GB	Effectively lossless. Use when disk and RAM are not the constraint.
Muse-Glimmer-30B-Q6_K.gguf	21.3 GB	Near-lossless; the last stop before quality loss becomes measurable.
Muse-Glimmer-30B-Q5_1.gguf	19.56 GB	Legacy. Prefer Q5_K_M.
Muse-Glimmer-30B-Q5_K_M.gguf	18.45 GB	Very good quality, noticeably smaller than Q6_K.
Muse-Glimmer-30B-Q5_K_S.gguf	18.02 GB	Slightly smaller than Q5_K_M, slight quality cost.
Muse-Glimmer-30B-Q5_0.gguf	18.02 GB	Legacy. Prefer Q5_K_M.
Muse-Glimmer-30B-Q4_1.gguf	16.47 GB	Legacy. Prefer Q4_K_M.
Muse-Glimmer-30B-Q4_K_M.gguf	15.77 GB	The usual default. Best quality-per-byte for most people.
Muse-Glimmer-30B-IQ4_NL.gguf	15.04 GB	Non-linear 4-bit; good on hardware without fast K-quant kernels.
Muse-Glimmer-30B-Q4_K_S.gguf	15.03 GB	A little smaller than Q4_K_M, a little worse.
Muse-Glimmer-30B-Q4_0.gguf	14.93 GB	Legacy round-to-nearest. Prefer Q4_K_M unless a runtime needs this.
Muse-Glimmer-30B-IQ4_XS.gguf	14.29 GB	Best sub-4.5 bpw option; usually beats Q4_K_S at a smaller size.
Muse-Glimmer-30B-Q3_K_L.gguf	13.67 GB	Small, with real quality loss. Usable when RAM is tight.
Muse-Glimmer-30B-Q3_K_M.gguf	12.74 GB	Smaller again; noticeable degradation.
Muse-Glimmer-30B-IQ3_M.gguf	11.94 GB	Strong at ~3.7 bpw, clearly better than Q3_K_M.
Muse-Glimmer-30B-IQ3_S.gguf	11.66 GB	Slightly smaller than IQ3_M.
Muse-Glimmer-30B-Q3_K_S.gguf	11.65 GB	Aggressive. Prefer IQ3_M at a similar size.
Muse-Glimmer-30B-IQ3_XS.gguf	11.15 GB	Aggressive but coherent.
Muse-Glimmer-30B-IQ3_XXS.gguf	10.38 GB	Very aggressive; the last coherent step down.
Muse-Glimmer-30B-Q2_K.gguf	9.95 GB	Very small, heavily degraded. For experimentation.
Muse-Glimmer-30B-Q2_K_S.gguf	9.34 GB	Smaller than Q2_K, at a further quality cost.
Muse-Glimmer-30B-IQ2_M.gguf	9.17 GB	The smallest size most people find usable.
Muse-Glimmer-30B-IQ2_S.gguf	8.5 GB	Below the usual usability line.
Muse-Glimmer-30B-Q2_0.gguf	8.34 GB	Extreme, group-64. Included for completeness.
Muse-Glimmer-30B-IQ2_XS.gguf	8.12 GB	Experimental.
Muse-Glimmer-30B-IQ2_XXS.gguf	7.41 GB	Experimental.
Muse-Glimmer-30B-IQ1_M.gguf	6.58 GB	Extreme. Expect substantial degradation.
Muse-Glimmer-30B-IQ1_S.gguf	6.08 GB	Extreme. Expect substantial degradation.
Muse-Glimmer-30B-Q1_0.gguf	4.52 GB	Extreme. Included for completeness.

32 files, ~483 GB total.

Running

bash

# one-off prompt
llama-cli -hf NANI-Nithin/Muse-Glimmer-30B-GGUF:Q4_K_M -p "Hello"

# OpenAI-compatible server
llama-server -hf NANI-Nithin/Muse-Glimmer-30B-GGUF:Q4_K_M

# download just one file
huggingface-cli download NANI-Nithin/Muse-Glimmer-30B-GGUF Muse-Glimmer-30B-Q4_K_M.gguf --local-dir .

# run a local file
llama-cli -m Muse-Glimmer-30B-Q4_K_M.gguf -p "Hello"

For image understanding, also download mmproj-Muse-Glimmer-30B-F16.gguf and load it as the projector with --mmproj.

Recommended settings (from Meta)

temperature = 1.0, top_p = 0.95, top_k = 64
Reasoning strength in the system prompt: Reasoning strength: <low | medium | high | xhigh>. Use high or xhigh for complex problem solving, coding, and agentic tasks.

Architecture

From the source model card:


Total parameters	~29.6B (including vision encoder)
Architecture	Dense causal transformer + perception encoder
Layers	52
Hidden dimension	6656
Attention	32 Q / 2 KV heads (GQA 16:1), head dim 128; [Local, Local, Local, Global] repeating, sliding window 2048
FFN	SwiGLU, intermediate dim 19,968
Position encoding	RoPE (θ = 500,000), local layers only
Vocabulary	202,048 (200k BPE + 2,048 special tokens)
Context length	131,072+
Vision encoder	~1.8B ViT-G/14, 50 layers, width 1536, patch 14 (arXiv:2504.13181)
Max visual tokens per image	4,096
Knowledge cutoff	January 4, 2026

Limitations

From the source card:

May produce inaccurate, biased, or objectionable responses; may still err in multi-step reasoning, especially in novel scenarios.
Not explicitly optimized for video (video is processed as individual frames). Audio input/output is not supported.
Quantized inference may show minor quality differences in edge cases compared to full precision — applies to every file in this repo.
Not intended to be downloaded by or used by individuals under 18.
Meta recommends deploying the model as part of a system with additional guardrails — in particular human-in-the-loop confirmation for irreversible actions in agentic contexts — rather than as an endpoint in itself.

Licence

Apache 2.0, same as the source model. Meta's usage policy is linked from the source repo.