Muse Glimmer: Meta's New Local Model for Always-On AI Agents
Meta has released the open weights of Muse Glimmer, a 30-billion-parameter model designed for always-on local AI agents. It runs on a single consumer GPU and, according to the company, stays close to frontier performance on agentic tasks thanks to distillation from much larger models.
What is Muse Glimmer
Muse Glimmer arrived on August 10, 2026 from Meta Superintelligence Labs under the Apache 2.0 license. It is a dense model: all 30B parameters are active on every inference, unlike mixture-of-experts models that only activate a subset. It is optimized for use cases such as local agents, function calling, assisted coding and LLM-as-a-judge evaluation.
Meta's bet is clear: bring AI to devices so it works without internet access or cloud infrastructure, anywhere, anytime.
How it was trained
Instead of pre-training on raw internet data, the model was trained with logit distillation on the outputs of Muse Spark, the larger model in the same family. It then went through a mid-training phase with long-context, agent-heavy data, and post-training combining supervised fine-tuning, on-policy distillation and reinforcement learning across general, reasoning, coding and agentic domains.
Built for agents
End-to-end task completion
The model is evaluated on full-task benchmarks such as DeepSearch QA, MCP-Atlas, tau-Bench and SWE-Bench, which measure its ability to work inside scaffolds, write and debug code, and resolve multi-turn requests from start to finish.
Failure recovery
One of its signatures: when a tool call fails or returns something unexpected, the model diagnoses the error and retries instead of halting. Add to that long-horizon multi-step reasoning, multimodal input (interleaved text and images), compatibility with orchestrators like OpenClaw, and training on more than 100 languages.
Optimized for local hardware
At full precision, 30B parameters would need more than 55 GB of memory, too much for a consumer GPU. Meta uses quantization to around 4-bit precision, shrinking the language model to under 20 GB and leaving room for the KV cache, the perception encoder and the speculative decoding drafter within a 24 or 32 GB envelope.
Speculative decoding
To gain speed, Muse Glimmer ships with a lightweight DFlash-based drafter that proposes entire blocks of tokens; the main model verifies them in parallel and corrects the wrong ones. The result is noticeably faster generation with identical output quality, according to Meta's measurements on MacBook M4-Max, M5-Max and an RTX 5090.
Availability
Weights have been on Hugging Face since day one. In the coming days integrations arrive for Ollama, LM Studio, Unsloth, llama.cpp, ExecuTorch, MLX, vLLM and SGLang, plus services such as Together AI, Fireworks AI and OpenRouter. The 4-bit quantization runs on GPUs like the RTX 3090, RTX 4090 or AMD 9700.
The competition does not rest
Meta positioned Muse Glimmer against Qwen3.6-27B and Gemma4-31B, which it beats on several tests. But four days later, Alibaba responded with Qwen3.8-27B, which tops Muse Glimmer on every benchmark where both were evaluated, such as Terminal Bench 2.1 (73.0 vs 51.7).
Conclusion
Muse Glimmer is an important step toward local agents: an open model that fits on a consumer GPU, understands images, recovers from errors and runs long tasks without depending on the cloud. The race for the best local model has just begun, and the competition between Meta and Alibaba is benefiting developers.
