Meta open-sources Muse Glimmer: 30B local agentic AI for consumer GPUs

Meta AI Research has open-sourced Muse Glimmer, a 30-billion-parameter model designed for always-on, local agentic workflows. Released under Apache 2.0, it lets developers run autonomous agents, complex tool invocation, local coding, and LLM-as-a-judge evaluations directly on consumer GPUs and workstations. To achieve this within strict memory budgets, Meta used a multi-stage training strategy distilled from its larger Muse Spark model: logit distillation for foundational reasoning, mid-training on long-context sequences with reasoning traces and tool call trajectories, and post-training alignment combining SFT, on-policy distillation, and reinforcement learning. A dedicated 1.8B parameter perception encoder enables native interleaved multimodal input, so agents can interpret screenshots, diagrams, and documentation inline during execution.

Runtimes are handled via two optimizations. Dynamic 4-bit quantization (K-Quant) shrinks the model footprint to roughly 17–20GB, leaving headroom in typical 24–32GB GPU/NPU setups for KV cache, perception embeddings, and speculative decoding. The DFlash speculative decoding pair uses a lightweight drafter to propose multi-token blocks, validated in parallel, boosting generation throughput up to 3.1x on Apple Silicon M4/M5 Max and NVIDIA RTX 5090. Muse Glimmer also handles long-horizon plans and failure states: when an API call or command errors, it diagnoses and retries alternate paths. It supports agent frameworks like OpenClaw, offers adjustable reasoning effort, and beats peers such as Gemma 4 31B and Qwen 3.6 27B on SWE-Bench, τ-Bench, and MCP-Atlas. Weights are on Hugging Face, with native support in llama.cpp, ExecuTorch, Apple MLX, Ollama, LM Studio, vLLM, and TorchTitan.