Meta Muse Glimmer 30B Brings Local AI Agents to Consumer Hardware
Muse Glimmer 30B is Meta's new open-weight model for local coding, tool use, and image understanding. Meta provides Apache 2.0 weights plus official GGUF and ExecuTorch artifacts. Its smallest GGUF is 16.8 GB before the optional vision encoder and speculative drafter, so the realistic target is a high-memory computer rather than a typical phone.
On this page
Meta released a 30B model for local agents
Meta released Muse Glimmer 30B on 10 August 2026. It is a dense, 29.6-billion-parameter model distilled from Muse Spark and released under the Apache 2.0 licence. Meta describes it as a model for coding agents, tool calls, multi-step tasks, failure recovery, and text-and-image input.
The release is more usable than a bare checkpoint. Meta published full-precision weights, two official GGUF builds for llama.cpp, and packaged ExecuTorch artifacts. The model has a 131,072-token context limit and a separate 1.8-billion-parameter perception encoder for images. Audio output and native video understanding are not supported.
That combination makes Muse Glimmer notable for local AI developers. A model can be downloadable yet remain awkward to deploy if it needs an unofficial conversion, an unmerged runtime change, or undocumented prompting. Meta coordinated the release with llama.cpp support and documented the required reasoning controls, chat format, and companion files.
The 17 GB download is only the starting point
The smaller official GGUF text model is 16.8 GB. Meta estimates roughly 17 GB of memory for text use, about 19 GB after adding the 1.4 GB vision encoder, and around 20 GB with the 1.6 GB speculative-decoding drafter as well. The higher-quality dynamic build raises those rough totals to 20 GB, 22 GB, and 23 GB.
Those figures explain Meta's 24 GB and 32 GB hardware targets. They leave some room for the context cache and runtime, but they do not mean every computer with that amount of advertised memory will behave the same. A discrete GPU's VRAM, unified memory on an Apple silicon Mac, system RAM with partial GPU offload, and memory already used by other applications are different constraints. Model size is only one part of the memory decision.
Long context also consumes memory as it fills. A 131K limit describes what the architecture supports, not a promise that the full window will remain fast or fit beside every combination of vision and speculative decoding. Most current Android phones do not have enough usable memory for these official builds. Muse Glimmer belongs in the desktop-class local model category unless smaller, properly tested variants arrive.
Meta's speed results need the stated hardware
Meta reports 233.4 tokens per second on an RTX 5090 when its DFlash drafter is enabled, compared with 74.9 without speculation. It reports 37.8 tokens per second on an M4 Max and 50.2 on an M5 Max, using ExecuTorch and the 17 GB quantised model. These are vendor measurements with batch size one and greedy decoding across Meta's prompt set.
The results show that speculative decoding can matter, but they are not general laptop benchmarks. DFlash proposes blocks of 16 tokens and lets the main model verify them in parallel. Its gain depends on how often those proposals are accepted, which can change with the prompt, sampling settings, and output. The drafter also consumes additional memory.
Meta says its 17 GB quantisation reduced average accuracy by one per cent across 15 benchmarks, while the dynamic build reduced it by 0.2 per cent. Those averages can hide larger changes on individual tasks. Developers should compare output quality, cold-load time, prompt processing, generation, heat, and peak memory on the machine they intend to support. Quantisation changes both fit and output quality, so the smallest file is not automatically the best choice.
Local execution does not make tool use safe
Keeping model inference on a computer can prevent prompts from being sent to a hosted model service. An agent can still use network tools, modify files, run commands, or disclose data through the actions connected to it. The model card recommends human confirmation for irreversible actions and application-specific safety tests.
Runtime versioning is another immediate check. The official GGUF files require llama.cpp build b10353 or newer; earlier builds do not recognise the Muse Glimmer architecture. Image input also needs the separate perception file. A download button should therefore identify the complete artifact set, check available memory, verify every file, and report an incompatible runtime before attempting a load.
For local AI users, the useful part of this release is not the broad claim that a 30B model runs on consumer hardware. It is the unusually complete path from licensed weights to official 17 GB and 20 GB runtime artifacts. That path is credible for well-equipped PCs and Macs, while phones and lower-memory computers remain outside the evidence Meta published.