Meta’s Muse Glimmer puts a 30B multimodal agent model on 24GB-class hardware

Meta released Muse Glimmer 30B in August as a 30-billion-parameter multimodal causal model distilled from Muse Spark and trained for autonomous agentic tasks. The official Hugging Face repositories include base weights, GGUF quantizations for llama.cpp and ExecuTorch artifacts for NVIDIA CUDA and Apple Silicon. Meta’s GGUF repository includes a roughly 17GB quantized build described as fitting comfortably in 24GB of VRAM, plus separate perception-encoder and DFlash speculative-decoding files. The repositories carry Apache 2.0 licensing metadata, while Meta also publishes a separate Muse Glimmer Usage Policy with prohibited-use restrictions.

A 30B agent model that can run on a single 24GB-class GPU changes the deployment floor for private coding, tool-use and multimodal agents. Builders can evaluate a capable local model without provisioning the multi-GPU infrastructure required by much larger open releases such as DeepSeek V4 Flash. The caveat is governance: the model repository is labeled Apache 2.0, but Meta’s separate usage policy imposes behavioral restrictions, so teams should not equate downloadable Apache-licensed weights with an absence of additional use conditions without reviewing the terms that apply to their deployment.

Muse Glimmer is not only a base-weight release. Meta publishes GGUF variants for llama.cpp and ExecuTorch PTE artifacts for CUDA and Apple Silicon, which lowers integration work for builders evaluating local inference. The official GGUF page says the 17GB K-quant build fits comfortably in 24GB of VRAM; image input requires the companion perception encoder.

Meta positions the model around multi-step reasoning, tool use, multimodal understanding, failure recovery and end-to-end agent task completion. The model card includes vendor-reported results on DeepSearch QA, MCP-Atlas, τ3-Bench and SWE-Bench. Those scores are useful directional evidence, but agent outcomes remain highly sensitive to harness, tool schema, token budget and retry policy.

Meta ships a DFlash drafter alongside the quantized model and documents speculative decoding support in local runtimes. This is relevant because local-agent usability depends as much on wall-clock latency as on benchmark quality. Builders should measure complete tool loops rather than raw token throughput alone.

The model repositories identify Apache 2.0 as the license, but Meta separately publishes a Usage Policy prohibiting several classes of use. BTN is not resolving the legal interaction between those documents. Commercial teams should review both rather than treating the repository license badge as the whole governance picture.