Muse Glimmer Runs a 30B Agentic Model on One Consumer GPU

An Open Agentic Model Sized for the Hardware You Already Own

Meta Superintelligence Labs released Muse Glimmer on August 10, 2026, and the headline number is not the parameter count but the memory footprint. At 30 billion parameters the model would normally demand more than 55GB of memory in full precision. Quantized to 4 bits it compresses to under 20GB, which puts it inside the envelope of a single consumer graphics card with 24GB to 32GB of memory. Meta says it tested the model on MacBook M4-Max and M5-Max laptops and on an RTX 5090 desktop card.

• 30 billion parameters, released under the permissive Apache 2.0 license on Hugging Face
• Under 20GB at 4-bit quantization, down from 55GB-plus in full precision
• 131K context window and coverage of more than 100 languages
• Native image understanding through a dedicated 1.8B-parameter perception encoder

That combination is what makes this release interesting rather than merely another entry in a crowded field. An open-weight model becomes genuinely useful the moment it fits on hardware people already have, and Muse Glimmer was clearly engineered backwards from that constraint.

What Does Muse Glimmer Actually Do Well?

Meta positions the model for agent workloads specifically: end-to-end task completion, tool invocation against precise schemas, multi-step reasoning chains, and recovery when a step fails. Those four capabilities usually live in separate places in a stack, with a large cloud model doing the reasoning and smaller local models handling narrow subtasks. Folding them into one 30B model that runs on a laptop changes where the boundary sits.

The multimodal side is not an afterthought. A dedicated perception encoder gives the model native image understanding, which Meta suggests for tasks like building a webpage from a design mockup or processing scanned documents. Reasoning strength is controllable across four settings — low, medium, high, and xhigh — so the same weights can serve a fast autocomplete loop and a slow planning pass.

On published benchmarks covering agentic behaviour, coding, multimodal understanding, reasoning, and safety, Meta describes the model as performing strongly for its size class against Gemma4-31B and Qwen3.6-27B. Treat cross-model comparisons from any vendor as a starting point rather than a verdict, and run your own evaluation on your own workload.

How Was It Trained?

The training recipe is worth noting because it explains the size-to-capability ratio. Meta used logit distillation from a larger teacher model during pre-training, extended context training during the mid-training phase, and then supervised fine-tuning that mixed on-policy distillation with reinforcement learning in post-training. Distillation from a stronger teacher is the standard way to get a small model punching above its weight, and Meta leaned on it at every stage rather than only at the end.

That approach also connects to the broader shift we covered in our guide to LLM quantization formats — the practical ceiling on local AI is memory bandwidth and capacity, not raw compute, so anything that shrinks the model without gutting it is worth more than an extra few points on a leaderboard.

Running It Today

Ollama shipped support the same day, initially through its MLX engine on Apple Silicon, where DFlash support delivers a claimed 1.5x to 1.8x speedup. Broader optimization for NVIDIA, AMD, and other platforms is listed as coming soon. On an RTX 5090, speculative decoding reportedly reaches a 3.1x speedup over the baseline. Integrations for llama.cpp, MLX, and ExecuTorch are available alongside the Hugging Face weights.

If you have not set up a local serving stack yet, our comparison of Ollama, vLLM, and llama.cpp walks through which runtime suits which workload. For readers tracking the smaller end of this trend, LFM2.5-2.6B running tool-calling agents in 2.5GB of RAM shows how far down the size curve capable agents now reach.

Why an Apache 2.0 License Matters Here

Apache 2.0 is about as permissive as open-weight licensing gets. It allows commercial use, modification, and redistribution without a separate agreement, which removes the legal review step that slows adoption inside companies. Meta previously shipped Muse Spark as a closed-weight frontier model and paired Muse Code with Muse Spark 1.2 earlier this month, so releasing the agentic model openly marks a deliberate split between the frontier tier and the local tier.

For developers, the practical upshot is simple. A capable agentic model with image understanding, tool calling, and a 131K context now runs offline, on one card, with no per-token cost and no network dependency. More local AI coverage lives in our artificial intelligence section.

Sources: Meta AI Research — August 10, 2026; Ollama Blog — August 10, 2026; MarkTechPost — August 10, 2026.