GenAIWiki

Muse Glimmer 30B

CurrentLatest

Muse Glimmer 30B is Meta Superintelligence Labs' open-weight multimodal model for local agents, coding, tool use, long-horizon reasoning, and image understanding.

Provider

Meta

Model family

Meta Muse

Open-weight multimodal agentic LLM

Cost tier

Glimmer 30b

Status

Current

Release Aug 10, 2026

Why teams choose it

🧠

The released model is Muse Glimmer 30B; Meta says open weights for a Muse Spark 1.2 vers…

ion are planned for the coming weeks, not available through this release.

📎

The full model accepts interleaved text and images

while the standalone quantized text-model GGUF needs the separate mmproj artifact for image input.

⚙️

Reasoning strength is controllable through low

medium, high, and xhigh settings in the system prompt.

Tradeoffs to know

  • Meta's benchmark results are vendor-reported and should be reproduced with the intended scaffold, runtime, quantization, and reasoning setting.
  • Agentic deployments require application-level permissions, prompt-injection defenses, validation, and human confirmation for irreversible actions.
  • Audio input and output are not supported; video is processed as frames and is not an explicitly optimized modality.
  • Runtime support for the new Muse Glimmer architecture, perception projector, and DFlash drafter may lag the weight release.

When not to use this

  • Self-hosting outcomes depend on hardware, quantization, and ops maturity—budget time beyond swapping an API hostname.
  • May demand more instrumentation than SaaS-managed APIs to duplicate latency, failover, and support guarantees.
  • Benchmark prompts and regressions continuously before rewriting entire routing tables around weights.

Technical specs

Inputs
text, image
Outputs
text
Capabilities
agentic task completion, coding, tool use, multimodal reasoning, failure recovery, controllable reasoning effort, multilingual generation, local deployment
License
Apache 2.0
Model string
muse-glimmer-30b

Benchmarks

{
  "source": "https://huggingface.co/meta-models/Muse-Glimmer-30B",
  "mmmu_pro": 74,
  "aime_2026": 94.7,
  "beam_128k": 65.1,
  "deepsearch_qa": 74.6,
  "swe_bench_pro": 51.2,
  "vendor_reported": true,
  "mcp_atlas_public": 75.5,
  "swe_bench_verified": 76
}

What Meta announced on August 10, 2026

Meta released the Muse Glimmer weights under Apache 2.0 as part of Mark Zuckerberg's announcement that the company is resuming open model releases. The announcement separately says Meta plans to open weights for a version of Muse Spark 1.2 in the coming weeks; that future Spark release is not the Muse Glimmer checkpoint available today.

  • Muse Glimmer 30B full-precision and quantized artifacts are publicly downloadable now.
  • Treat Muse Spark 1.2 open weights as announced but not yet released until Meta publishes the checkpoint and license.
  • Use the exact model name in evaluations and telemetry because Glimmer and Spark target different capability and deployment lanes.

Sources: Meta Muse Glimmer model card, Axios interview and announcement summary

Muse Glimmer 30B architecture and modalities

Meta documents Muse Glimmer as a dense causal transformer with approximately 29.6 billion parameters including a dedicated roughly 1.8B-parameter ViT-G/14 perception encoder. It accepts text and images, generates text, and has a documented context length of 131,072+ tokens.

  • The language model uses 52 layers and a repeating local-local-local-global attention pattern.
  • The model card lists a January 4, 2026 knowledge cutoff and training across more than 100 languages.
  • The perception path supports screenshots, charts, documents, and images; audio is not supported.

Sources: Muse Glimmer architecture and intended use

Muse Glimmer local deployment options

Meta publishes BF16 weights plus two GGUF quantizations. The official model card targets the full-precision build at 64 GB VRAM, the dynamic K-quant build at 32 GB, and the 17 GB K-quant build at 24 GB. These are Meta's tested target envelopes, not universal minimum requirements.

  • The GGUF text model runs without vision; add mmproj-kquant.gguf for image input.
  • The optional dflash-kquant.gguf drafter enables speculative decoding only when the selected runtime supports the released DFlash integration.
  • Start with a smaller context than 131K and measure KV-cache memory, prompt-processing time, generation speed, and task quality on the actual device.

Sources: Official Muse Glimmer GGUF artifacts

Agent safety and evaluation checklist

A locally running model can still create external risk when connected to files, browsers, shells, credentials, or business systems. Evaluate the complete agent scaffold rather than treating the base model's benchmark table as a deployment approval.

  • Validate tool arguments and restrict each tool to the minimum identity, filesystem, network, and data permissions required.
  • Require confirmation for purchases, messages, account changes, production writes, secrets access, and destructive actions.
  • Test prompt injection, recovery from failed tools, long-horizon drift, multilingual quality, privacy boundaries, and quantization regressions.

Sources: Muse Glimmer trust, safety, and limitations, Meta evaluation methodology


Meta Muse family lineup


Compare with

Muse Glimmer 30B FAQ

What is Muse Glimmer 30B?

Muse Glimmer 30B is Meta Superintelligence Labs' open-weight multimodal model for local agents, coding, tool use, long-horizon reasoning, and image understanding. Meta's model card documents a dense 29.6B-parameter architecture with a dedicated perception encoder, 131,072+ context, text-and-image input, text output, c...

When does Muse Glimmer 30B fit best?

Local coding and research agents

What should teams watch out for with Muse Glimmer 30B?

Meta's benchmark results are vendor-reported and should be reproduced with the intended scaffold, runtime, quantization, and reasoning setting.

Explore next

Models, tools, and comparisons that connect to this reference.