Meta released Muse Glimmer on August 10, a 30-billion-parameter open-weight model built specifically to run always-on agentic workloads locally, on a single consumer GPU or a Mac, with no server round-trip required. The model comes from Meta Superintelligence Lab and is distilled from Muse Spark, transferring reasoning capability from Meta’s larger flagship model down into a package small enough to sit on a developer’s own machine.
Built for local agent loops, not chat
Meta compresses Muse Glimmer to roughly 4-bit precision and pairs it with a lightweight companion model for speculative decoding, a technique that lets a smaller draft model propose tokens a larger model can quickly verify, speeding up generation enough for the model to function inside a real-time agent loop rather than just answering one-off prompts. Meta positions the release around local agents, function calling, on-device coding assistance, and using the model as an evaluator (“LLM-as-a-judge”) for other local workflows. Weights are released under the permissive Apache 2.0 license, and Meta says integrations with llama.cpp, MLX, and ExecuTorch are coming in the following days, alongside support from Ollama, LM Studio, Together AI, and Fireworks AI.
Zuckerberg pledges to open Muse Spark 1.2
Meta CEO Mark Zuckerberg tied the release to a broader open-source push, saying Meta plans to release open weights for Muse Spark 1.2 — the flagship model Glimmer was distilled from — in a future update. If Meta follows through, it would put a genuine frontier-tier model into open circulation from a major U.S. lab, a step neither OpenAI, Anthropic, nor Google has taken with their current top-tier systems.
Why it matters
Muse Glimmer arrives days after Meta’s own Muse Code terminal coding agent and amid a broader wave of agent-focused releases from OpenAI, xAI, and Anthropic. Its significance is less about benchmark scores than deployment economics: a capable model that runs entirely on local hardware removes the latency, cost, and data-residency concerns that come with routing every agent action through a cloud API, which could make locally hosted agents more viable for developers and enterprises wary of sending code or business data to a third-party model provider. It also puts renewed competitive pressure on other open-weight labs — including Mistral, Alibaba’s Qwen team, and Moonshot AI — whose smaller models have been the default choice for on-device and edge agent deployments.