Meta Open-Sources Muse Glimmer: A 30B Local Agentic Model Optimised for On-Device Execution

Details

Meta AI Research has released Muse Glimmer, a 30B open-weight model licensed under Apache 2.0 and designed for always-on local agentic workflows. The model targets developers who need autonomous agents, complex tool invocation, local coding assistance, and LLM-as-a-judge evaluation without depending on cloud APIs. It is distilled from Meta's larger flagship, Muse Spark. Training followed a three-stage pipeline: logit distillation from Muse Spark for foundational reasoning transfer, mid-training on long-context sequences with reasoning traces and multimodal interleaved data, then post-training via Supervised Fine-Tuning, on-policy distillation, and Reinforcement Learning for code generation, tool usage, and structured planning. A dedicated 1.8B perception encoder processes interleaved multimodal inputs -- screenshots, diagrams, documentation -- natively during agent execution. To fit consumer hardware, Muse Glimmer uses 4-bit dynamic K-Quant compression, dropping the uncompressed 55GB+ footprint to 17-20GB -- usable within standard 24-32GB GPU/NPU envelopes with headroom for KV cache and speculative decoding overhead. DFlash speculative decoding pairs the base model with a lightweight companion drafter that proposes multi-token blocks validated in parallel, yielding up to 3.1x throughput on Apple Silicon M4/M5 Max and NVIDIA RTX 5090. The model is trained to execute long-horizon plans and handle failure states: when an API call or terminal command returns an error, it diagnoses the failure and attempts recovery. Apache 2.0 licensing makes commercial deployment straightforward, distinguishing it from many gated open-weight releases.