Meta has released Muse Glimmer, a 30-billion-parameter open-weight model built to run agentic AI on a single computer, publishing the weights and the official model card on Hugging Face under the Apache 2.0 license. The release extends Meta’s open-weights line into the consumer hardware tier, where Meta Superintelligence Labs says the model can handle multi-step agent tasks without cloud infrastructure.
Muse Glimmer is distilled from Muse Spark, the agentic model family Meta launched earlier this year, and pairs a dense causal transformer with a perception encoder that accepts images alongside text. An NVIDIA technical blog post published the same day details how the model runs locally on NVIDIA hardware, quoting 20,000 tokens per second on a single GPU and listing GeForce RTX 5090, DGX Spark, DGX Station and Jetson platforms, with deployment options through NVIDIA NIM containers, SGLang and vLLM.
Specs aimed at a 24 GB envelope
The model card lists roughly 29.6 billion parameters, including a 1.8-billion-parameter ViT-G/14 vision encoder, 52 layers with a 32:2 grouped-query attention pattern, SwiGLU feedforward blocks and RoPE position encoding. Context length is 131,072 tokens, with up to 4,096 visual tokens per image, and the tokenizer covers 200,000 BPE tokens. The knowledge cutoff is January 4, 2026, and training data spans more than 100 languages.
The local deployment story is the headline. Meta says 4-bit quantization compresses the language model to under 20 GB, which leaves room for the KV cache, the perception encoder and a speculative decoding drafter inside a 24 GB or 32 GB memory envelope. The company measured 0.2 percent accuracy degradation on 32 GB targets and 1.0 percent on 24 GB targets across an average of 15 common benchmarks, arguing the compression costs almost nothing on agentic workloads. That is the same argument this site has tracked in the case for running local AI models on consumer hardware.
Muse Glimmer is benchmarked on agentic tasks rather than chat: DeepSearch QA, MCP-Atlas, tau3-Bench and SWE-Bench, covering tool use, coding and multi-turn task completion. NVIDIA frames the design choice explicitly, noting the dense architecture activates every parameter per token, with no mixture-of-experts routing, which buys reliability, long-context coherence and predictable latency for long-running agents. The model card adds failure recovery, letting the model diagnose failed tool calls and retry, and controllable effort settings that trade quality against speed.
The release lands in the middle of the open-weights debate this site has covered before, in the argument that open-weight models are making their hardware case. A 30B model that fits in a gaming GPU’s memory does not need a rented cluster for everyday agent work, which changes the hardware conversation from scale to locality: private inference, offline operation and no per-token cloud costs.
Downloads are live now on Hugging Face, and the model runs through the usual serving stacks on day one. The practical question is whether a dense 30B model at 4-bit precision can hold up in real agent scaffolds over long sessions; the benchmarks say yes on paper, and the weights are out there for anyone to check.
