Muse Glimmer 30B

Best for

Builders running persistent local agents that need reliable tool calling, self-managed memory, and recovery from failed steps, on a 24 GB GPU or a 32 GB Mac. Also worth a look for image-aware agent workflows, since it accepts interleaved text and images natively.

Who should use it

• Builders running persistent local agents that need reliable tool calling, self-managed memory, and recovery from failed steps, on a 24 GB GPU or a 32 GB Mac. Also worth a look for image-aware agent workflows, since it accepts interleaved text and images natively.
• Builders who want local or self-hosted testing options.
• Teams testing tool-use, agentic planning, and multi-step workflow behavior.

Common workflows

• Always-on local agents with tool calling, persistent memory, and image input
• open weights workflows
• local workflows
• agents workflows
• multimodal workflows

Deployment and hardware notes

Dense 29.8B parameters (a 27.9B language model plus a 1.92B vision projector). Ollama's default Q4_K_M is about 18 GB, closer to 19.5 GB once runtime overhead is counted: comfortable on a 24 GB GPU such as an RTX 3090 or 4090, with room for KV cache and the perception encoder. Q8_0 is about 31 GB and BF16 about 57 GB (Ollama tag) to 64 GB (Meta's figure), both server or Mac Studio class. A 32 GB unified-memory Mac fits Q4 with headroom. Meta also publishes official K-Quant-Dynamic (32 GB target) and K-Quant-17GB (24 GB target) checkpoints on Hugging Face. No independent local throughput measurements were available at review time, so treat speed as something to test on your own hardware.

License and usage notes

Apache 2.0. Open weights. Verify the exact model card and license terms for the checkpoint or hosted provider you use.

Strengths

• Open weights model option for Muse workflows.
• Builders running persistent local agents that need reliable tool calling, self-managed memory, and recovery from failed steps, on a 24 GB GPU or a 32 GB Mac. Also worth a look for image-aware agent workflows, since it accepts interleaved text and images natively.
• Start with `ollama run muse-glimmer:30b` (18 GB Q4_K_M, 128K context, text and image input). Ollama also lists 30b-q8_0 (31 GB), 30b-bf16 (57 GB), and 30b-mlx (21 GB) tags. In LM Studio, pull lmstudio-community/Muse-Glimmer-30B-GGUF. llama.cpp needs release b10353 or newer; vLLM supported it on day one. Meta says it works across OpenClaw, Hermes Agent, and other agentic orchestration patterns, and ships a DFlash speculative-decoding drafter for faster generation where your runtime supports it.
• Tracked as Practical local in the OpenSourcesAI model directory.

Limitations

• The weights are Apache 2.0, but Meta ships a separate usage policy alongside the license, so read both before you deploy. Q4 needs 24 GB class hardware for a full-GPU fit; 12 to 16 GB cards run it only with CPU offload. Native MLX support upstream is still in progress. Agentic tool-calling quality is what the model is built for, so test it on your own tasks and tool schemas rather than relying on the announcement.
• Dense 29.8B parameters (a 27.9B language model plus a 1.92B vision projector). Ollama's default Q4_K_M is about 18 GB, closer to 19.5 GB once runtime overhead is counted: comfortable on a 24 GB GPU such as an RTX 3090 or 4090, with room for KV cache and the perception encoder. Q8_0 is about 31 GB and BF16 about 57 GB (Ollama tag) to 64 GB (Meta's figure), both server or Mac Studio class. A 32 GB unified-memory Mac fits Q4 with headroom. Meta also publishes official K-Quant-Dynamic (32 GB target) and K-Quant-17GB (24 GB target) checkpoints on Hugging Face. No independent local throughput measurements were available at review time, so treat speed as something to test on your own hardware.
• Context window and limits: 131,072 tokens.
• Verify the exact model card, provider docs, license, and serving support before production use.

Local workflow notes

Start with `ollama run muse-glimmer:30b` (18 GB Q4_K_M, 128K context, text and image input). Ollama also lists 30b-q8_0 (31 GB), 30b-bf16 (57 GB), and 30b-mlx (21 GB) tags. In LM Studio, pull lmstudio-community/Muse-Glimmer-30B-GGUF. llama.cpp needs release b10353 or newer; vLLM supported it on day one. Meta says it works across OpenClaw, Hermes Agent, and other agentic orchestration patterns, and ships a DFlash speculative-decoding drafter for faster generation where your runtime supports it.

Local runtimes: Ollama (muse-glimmer:30b), llama.cpp (b10353+), LM Studio, vLLM, MLX via Ollama 30b-mlx

Platforms: Windows, macOS, Linux

VRAM fit by quantization level

Enter your GPU VRAM below to see which quantization of Muse Glimmer 30B fits and get the Ollama run command.

Sources to verify

Additional sources

Related resources

Continue with model source notes, local tools, and implementation guides related to this model.

Ready to run this model locally?