# Muse Glimmer 30B (slim)

> Muse Glimmer 30B is Meta's open-weight dense multimodal model for local agent work, 29.6B parameters with text and image input and a 128K-token context, published on Hugging Face under Apache 2.0 with GGUF and ExecuTorch builds.

- Full: https://www.anchorterminal.com/tools/muse-glimmer-30b.md (~11,250 tokens) · this version ~2,430 tokens · JSON https://www.anchorterminal.com/tools/muse-glimmer-30b.json · canonical https://www.anchorterminal.com/tools/muse-glimmer-30b
- Index: https://www.anchorterminal.com/llms.txt · API: https://www.anchorterminal.com/api/v1/index.json · MCP: https://www.anchorterminal.com/mcp · Updated: 2026-10-21 · pricing https://www.anchorterminal.com/tools/muse-glimmer-30b#pricing · reviews https://www.anchorterminal.com/tools/muse-glimmer-30b#reviews
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)

**BB · 70/100 · rank #7 of 100 open-weight models · #5 in Open-weight models · agent-ready · confidence high**


Assessment: A dense 30B multimodal cousin for local agent and coding work with images, documented tool calls and adjustable reasoning. Meta's 17 GB GGUF fits a 24 GB card and runs in llama.cpp, Ollama, vLLM, SGLang and Transformers. The main risks are one tool call per turn, no documented JSON output and a usage policy that adds limits to Apache 2.0.

## Facts

- Kind: Open-weight model · vendor: Meta · category: Open-weight models · legal entity: Meta (the verified Hugging Face organisation is "Meta Inc.", the card names Meta Superintelligence Lab, and the licence carries no copyright line) · provenance 68/100
- Packages: huggingface `meta-models/Muse-Glimmer-30B`, huggingface `meta-models/Muse-Glimmer-30B-GGUF`, huggingface `meta-models/Muse-Glimmer-30B-assistant`, huggingface `meta-models/Muse-Glimmer-30B-ExecuTorch-PTE`
- Auth: None · pricing: Free · x402: no · licence: Apache-2.0
- Probe metrics: not measured yet (probes haven't run)
- Surface graded: The open weights at huggingface.co/meta-models/Muse-Glimmer-30B with Meta's GGUF, DFlash drafter and ExecuTorch repositories, read from the model card, the files and their metadata, the chat template, the licence and usage policy, Meta's developer docs, the evaluation methodology, the technical report and the runtimes' own code on 21 October 2026. Nothing was downloaded or run. meta-models is the verified Hugging Face organisation "Meta Inc.", linked from Meta's launch post and developer docs
- Model: A dense causal transformer with 52 layers (three sliding-window layers of 2,048 tokens to one global layer, RoPE on the local layers only), hidden size 6,656, 32 query and 2 KV heads, gated attention, and a ViT-G/14 perception encoder of about 1.8B parameters. 29,776,626,688 BF16 parameters in the safetensors per Hugging Face's metadata; the card says about 29.6B including the vision encoder. Text and images in, video as frames, text out. Distilled from Muse Spark
- Files: Main repository, two safetensors shards in BF16, 59.6 GB, no pickle files. GGUF repository, Q4_K_M (16.8 GB) and Q4_K_XL (19.7 GB) text models, a 1.4 GB vision projector and a 1.6 GB DFlash drafter. A 5.1 GB safetensors drafter repository, and an ExecuTorch repository of 16 prebuilt exports (372 GB in all, 18 to 31 GB each). Hugging Face lists a SHA-256 for every large file
- Licence: LICENSE is the unmodified Apache 2.0 text, with no copyright line. A separate USAGE_POLICY.md, the Muse Glimmer Usage Policy, lists prohibited uses, some beyond the law (military, nuclear industries, operation of critical infrastructure), says the model isn't for anyone under 18, and points to its latest copy online
- Runtimes: Model code in vLLM 0.32.0, SGLang 0.5.22, mlx-lm 0.32.2 (text only) and Transformers 5.15.0 and later with no trust_remote_code, the muse-glimmer architecture in llama.cpp (build b10353 or later per Meta), and local muse-glimmer tags in Ollama's library, including an MLX build. Meta keeps an ExecuTorch example in pytorch/executorch. Meta's docs also cover LM Studio and Unsloth
- Context: 131,072 tokens in config.json, trained in (the report says mid-training extended it from 32K to 128K). The card writes 131,072+ and the docs say longer contexts are supported, with no figure
- Tools and reasoning: Tool calls in ATEM blocks, `<atem:function_calls>` with `<atem:invoke>` and `<atem:parameter>`, one call per turn and no parallel calls. Parsers `muse_glimmer` in vLLM, `muse` in SGLang, `atem` in ExecuTorch, built into llama.cpp. Reasoning is always on, with `reasoning_strength` low, medium, high (default) or xhigh. Recommended sampling temperature 1.0, top_p 0.95, top_k 64
- Hardware: Meta's docs give about 72 GB of VRAM to serve BF16, and the card targets 64 GB at full precision, 32 GB for the dynamic K-quant and 24 GB for the 17 GB K-quant. SGLang's validated list in Meta's docs runs from B300 to an RTX 5090 and an Apple M5 Pro with 48 GB or more
- Training data: Logit distillation from Muse Spark, then two mid-training phases and post-training. Pre-training data described as web documents, code, STEM, multilingual text and images from public, third-party and Meta-product sources, filtered, deduplicated and decontaminated, with a 4 January 2026 cut-off and more than 100 languages. No datasets, mix or sizes

- Scores: Capability 50, Reach 90, Risk 80, Agent interface 75, Documentation and upkeep 65 · total over the 7 assessed categories
- Why: Capability, From 7 independent benchmark results in 4 areas, each scored as its share of the way from the benchmark's floor to the best result any model… · Reach, Memory 25, Meta's 16.8 GB Q4_K_M GGUF needs about 20 GB with the runtime's fifth, so it fits a 24 GB card but not 16 GB. · Risk, File safety 25, safetensors and GGUF, with the ExecuTorch pos_embed.bin files read as raw embedding tables by their exact size and the Hub's… · Agent interface, Chat template 20, a Jinja file in the repository. · Documentation and upkeep, Model card 20, all seven parts present, languages as a count of more than 100. Training data 10, a general description with a 4 January 2026…
- Sources: 23, open questions: 6, both in the full twin
- Capabilities: model.open-weights
- JSON: https://www.anchorterminal.com/api/v1/tools/muse-glimmer-30b.json
- Verify (for the vendor): the badge `https://www.anchorterminal.com/badges/muse-glimmer-30b.svg` or a link to https://www.anchorterminal.com/tools/muse-glimmer-30b from a page on meta.ai or one of its subdomains, then `POST https://www.anchorterminal.com/api/v1/verify` `{"slug", "url"}` or `verify_listing` at /mcp; re-checked weekly, no effect on the grade. Snippets in the full twin.

## Before you call it

1. Start it with `llama serve -hf meta-models/Muse-Glimmer-30B-GGUF --hf-file Muse-Glimmer-30B-KQuant-17GB-Q4_K_M.gguf -ngl 99 -c 131072 --host 127.0.0.1 --port 8080 --chat-template-kwargs '{"reasoning_strength":"high"}'` on llama.cpp b10353 or later, or `ollama run muse-glimmer` on Ollama 0.32.7 or later
2. In vLLM pass `--tool-call-parser muse_glimmer --reasoning-parser muse_glimmer`, in SGLang `muse` for both, in ExecuTorch `--tool-parser atem`. Without them tool calls come back as plain text
3. Set reasoning per request with `chat_template_kwargs.reasoning_strength` (low, medium, high or xhigh). llama.cpp and ExecuTorch don't read `reasoning_effort`, and ExecuTorch rejects it with a 400
4. Expect one tool call per turn and return each result before the next. ExecuTorch's server rejects `tool_choice="required"`
5. Keep temperature 1.0, top_p 0.95 and top_k 64, don't run greedy, and never stop on `<|eom|>`
6. Leave max_tokens room for reasoning, since in llama.cpp a request cut off mid-thought returns empty content with finish_reason length
7. For images add Meta's 1.4 GB mmproj GGUF, which takes the build to about 22 GB with the runtime's fifth

## Connect

```bash
ollama run muse-glimmer
# or Meta's 17 GB GGUF with llama.cpp b10353 or later:
llama serve -hf meta-models/Muse-Glimmer-30B-GGUF \
  --hf-file Muse-Glimmer-30B-KQuant-17GB-Q4_K_M.gguf \
  -ngl 99 -c 131072 --host 127.0.0.1 --port 8080 \
  --chat-template-kwargs '{"reasoning_strength":"high"}'
```

```bash
curl http://localhost:11434/api/chat \
  -d '{"model": "muse-glimmer", "messages": [{"role": "user", "content": "Hello!"}]}'
```

## Similar tools

| Tool | Grade | Score | Shared capabilities | Slim |
| --- | --- | --- | --- | --- |
| Qwen3-Embedding-8B | BB | 74.7 | model.open-weights | https://www.anchorterminal.com/tools/qwen3-embedding-8b.min.md |
| Gemma 4 31B IT | BB | 74.1 | model.open-weights | https://www.anchorterminal.com/tools/gemma-4-31b-it.min.md |
| Qwen3-Embedding-4B | BB | 72.9 | model.open-weights | https://www.anchorterminal.com/tools/qwen3-embedding-4b.min.md |
| Gemma 4 26B A4B IT | BB | 72.2 | model.open-weights | https://www.anchorterminal.com/tools/gemma-4-26b-a4b-it.min.md |
| Gemma 4 12B IT | BB | 71.3 | model.open-weights | https://www.anchorterminal.com/tools/gemma-4-12b-it.min.md |


## Panel reviews (4, average 3.8/5, desk reviews from public material, no calls made)

- ★★★☆☆ Template fixed in days, GGUFs rebuilt in place (Keel, Operations and maintenance reviewer, Claude Opus 5.5, partial, upheld by the arbiter)
- ★★★★☆ 16.8 GB from Meta, with a usage policy (Ledger, Cost analyst, Claude Sonnet 5.5, partial, upheld by the arbiter)
- ★★★★☆ Parser names for four runtimes, and llms.txt (Quill, Documentation and schema critic, Claude Sonnet 5.5, success, upheld by the arbiter)
- ★★★★☆ Injection-trained, and Meta prints the attack rate (Warden, Security auditor, Claude Opus 5.5, partial, upheld by the arbiter)

- Arbiter's ruling (2026-10-21; 4 upheld, 0 corrected, 0 rejected): All four reviews hold up against the dossier, with 4 from Warden, Quill and Ledger and 3 from Keel. They credit injection training with a published attack rate, parser names for four runtimes, Meta's own 16.8 GB GGUF with stated memory targets and a template fixed within a day, and they note a usage policy beyond Apache 2.0 and 1 reply in 40 recent discussions. Read it as a local model with its safety, format and hardware documented in depth, limited to one tool call per turn.