# Meta’s Muse Glimmer puts a 30B multimodal agent model on 24GB-class hardware Muse Glimmer is a 30B multimodal agent model released with Apache 2.0 weights, official GGUF and ExecuTorch artifacts, and a quantized build designed to fit in 24GB of VRAM — giving builders a materially more practical local-agent option, with an additional Meta usage policy that still needs governance review. Meta’s Muse Glimmer 30B combines tool use, coding, vision and agentic task completion with official local-runtime artifacts. A 17GB GGUF build targets 24GB-VRAM machines, but Meta also attaches a separate usage policy, so builders should distinguish weight availability from unrestricted use. - Status: Active - Published: 2026-08-22T10:57:17+12:00 - Updated: 2026-08-22T10:57:17+12:00 - Categories: Artificial Intelligence, AI Agents, Open Models, Inference & APIs - Tags: AI agents, local AI, Meta, Muse Glimmer, open weights - Canonical HTML: https://beyondthe.news/dossiers/meta-muse-glimmer-30b-local-agent-model ## What changed 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. ## Why it matters 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. ## The practical release is the local-runtime packaging 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. ## It is explicitly trained for agentic workloads 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. ## Speculative decoding is part of the deployment story 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. ## Apache 2.0 metadata does not remove policy review 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. ## Key details - Muse Glimmer is a 30B causal language model with a dedicated perception encoder and multimodal input support. - Meta’s official repositories identify Apache 2.0 as the model license and also publish a separate Muse Glimmer Usage Policy. - Official GGUF artifacts include a roughly 17GB quantized build described as fitting comfortably in 24GB of VRAM. - The GGUF path supports llama.cpp; Muse Glimmer support requires llama.cpp build b10353 or newer according to Meta’s repository. - Official ExecuTorch artifacts target NVIDIA CUDA and Apple Silicon Metal. - A companion DFlash model supports speculative decoding, while a separate perception encoder is needed for image input in the GGUF path. - Meta’s agent and coding benchmarks are vendor-reported and should be validated with workload-specific evaluations. ## Builder takeaways - If you want a private or offline agent, benchmark Muse Glimmer on a single 24GB-class GPU before assuming you need multi-GPU open-model infrastructure. - Use Meta’s official GGUF or ExecuTorch artifacts first; they reduce conversion risk and make it easier to reproduce the documented local path. - Measure end-to-end agent completion time with tools enabled. Speculative decoding may improve responsiveness, but tool latency and retry loops can dominate. - Re-run quality tests on the exact quantization you plan to deploy; vendor benchmark results for the base model do not automatically transfer to every compressed artifact. - Review both the Apache 2.0 repository license and Meta’s separate Usage Policy with your legal or governance owner before commercial deployment. ## What to watch - Independent agent and coding evaluations of Muse Glimmer using fixed public harnesses. - Real-world throughput and memory results for the official 17GB GGUF build on 24GB consumer GPUs and Apple Silicon. - Whether Meta clarifies how the separate Usage Policy interacts with the Apache 2.0 repository licensing for downstream commercial use. - Support maturity in llama.cpp, vLLM, ExecuTorch and other common local-serving stacks. - Whether Meta releases smaller Glimmer variants or updated distilled models that preserve agent performance at lower memory footprints. ## Uncertainties - Meta’s headline agent benchmarks are vendor-reported and agent-harness sensitive. - The legal relationship between the Apache 2.0 license metadata and Meta’s separate Usage Policy is not resolved here; builders should review the applicable terms for their own use case. - The phrase 'consumer hardware' spans widely different devices; the official 17GB GGUF target is concrete, but practical latency and context limits depend on hardware and runtime configuration. ## Sources - [Muse Glimmer 30B model card](https://huggingface.co/meta-models/Muse-Glimmer-30B) — Meta / Hugging Face · primary · 2026-08-01T00:00:00+12:00. Official model description, architecture, intended agentic capabilities, license metadata and vendor benchmarks. - [Muse Glimmer 30B GGUF](https://huggingface.co/meta-models/Muse-Glimmer-30B-GGUF) — Meta / Hugging Face · primary · 2026-08-01T00:00:00+12:00. Official quantized artifacts, 17GB/24GB-VRAM deployment guidance, llama.cpp requirement, perception encoder and DFlash files. - [Muse Glimmer Usage Policy](https://huggingface.co/meta-models/Muse-Glimmer-30B/blob/main/USAGE_POLICY.md) — Meta / Hugging Face · primary · 2026-08-01T00:00:00+12:00. Separate prohibited-use policy that builders should review alongside the repository license. - [Meta is back with Muse Glimmer: local, agentic, multimodal, and open source](https://huggingface.co/blog/muse-glimmer) — Hugging Face · ecosystem_analysis · 2026-08-10T00:00:00+12:00. Independent ecosystem walkthrough of local runtimes, speculative decoding, fine-tuning and agent usage.