Muse Glimmer

Muse Glimmer is Meta's open-weight 30B parameter agentic multimodal model, released on August 10, 2026 under the Apache 2.0 license. Distilled from Meta's larger Muse Spark model, it is specifically engineered for always-on local agent workflows: planning, tool calls, result checking, and failure recovery. It accepts interleaved text and images, runs on a single consumer GPU (24GB target with 4-bit quantization), and needs no network call, making it a credible default for private local coding agents.

Core Features

• Agent-native training: Built around the agent loop (plan, call tools, interpret results, recover from failure), not just chat.
• Multimodal input: A ~1.8B ViT-G/14 perception encoder accepts screenshots, charts, and documents alongside text, up to 4,096 visual tokens per image.
• Runs locally: Dense 30B model with ~29.6B total parameters, 131,072+ token context, Apache 2.0 license, no network required.
• 24GB target: 4-bit quantized builds fit the language model under 20GB, leaving room for KV cache, vision encoder, and drafter in a 24GB/32GB envelope.
• DFlash speculative decoding: A 2.56B block-diffusion drafter proposes 16 tokens per forward pass, verified in parallel; up to 3.1x decode speedup on RTX 5090.
• Day-0 ecosystem: Support in transformers, llama.cpp, vLLM, Ollama, LM Studio, SGLang, MLX, and ExecuTorch at launch.

Model Specifications

Benchmark Highlights

• BionicBench (LM Studio, independent): 83.3% task completion vs 77.7% for Gemma 4 31B and Qwen3.6-27B.
• Agentic wins: Leads comparators on MCP Atlas, DeepSearch QA, and SWE-Bench Pro.
• Tradeoffs: Trails Qwen3.6-27B on OSWorld-Verified, TerminalBench 2.1, and SkillsBench for long multi-hour sessions.

Use Cases

• Private local coding agents: Run tool-use agents fully offline with no data leaving the machine.
• Document and screen understanding: Agents that read screenshots, charts, and documents while calling tools.
• Personal assistants: Claw/Hermes-style setups that need long-horizon context and multimodal perception.
• Cost-sensitive pipelines: LLM-as-a-judge and repetitive evaluation workloads are cheap to run locally.

Advantages

1. Truly local agentic: A 30B multimodal agent that fits one consumer GPU, with a fully permissive license.
2. Integrated stack: Quantization, speculative decoding, and runtimes designed together rather than bolted on.
3. Meta ecosystem: Official GGUF, ExecuTorch, and transformers support plus named hardware partners (AMD, Arm, Dell, Intel, NVIDIA).

Tips

1. Start with K-Quant-17GB on a 24GB card: Validate text-only performance first, then add vision and DFlash while watching memory headroom.
2. Use DFlash in agent loops: Speculative decoding matters most for long reasoning chains and multi-step tool calls.
3. Pick Qwen3.6-27B for multi-hour terminal sessions: Glimmer shines at scoped tool use; its rivals lead on very long autonomous runs.

Conclusion

Muse Glimmer gives local-agent developers their first realistic default: a permissively licensed, multimodal, agentic 30B model that runs on hardware they already own. If you want private, always-on coding and tool-use agents without cloud dependence, it is the strongest open-weight option in its class today.