Meta has open-sourced Muse Glimmer, a 30-billion-parameter model distilled from its larger Muse Spark 1.2, designed to run local AI agents on consumer hardware with no cloud dependency. Released under Apache 2.0 with weights on Hugging Face, Glimmer fits on a single GPU with 24GB VRAM and is optimized for agentic workflows including function calling, code generation, file and screenshot manipulation, and multi-step task planning. The model supports multimodal input (text + images) and runs on consumer GPUs (RTX-series), Mac M4/M5 Max, and AMD Ryzen AI hardware.
Glimmer was trained using a three-phase pipeline: pre-training via logit distillation on Muse Spark's outputs; mid-training on extended-context, agent-heavy data with richer reasoning traces; and post-training combining supervised fine-tuning with on-policy distillation and reinforcement learning. A lightweight secondary DFlash-based drafter model uses speculative decoding, proposing blocks of 16 tokens for parallel verification. On NVIDIA RTX 5090, this increases generation speed from 74.9 to 233.4 tokens per second; on M4 Max from 23.7 to 37.8 tokens/sec; on M5 Max from 26.6 to 50.2 tokens/sec.
For development teams and enterprises building agentic applications, Glimmer removes cloud latency and per-token API costs from the loop. Benchmarks show leadership on MCP Atlas (75.5) and DeepSearch QA (74.6) versus Gemma4-31B and Qwen3.6-27B, signaling strong agentic reasoning at the edge. Meta frames this as part of Zuckerberg's 'personal superintelligence' vision—distributing capable AI to individual devices rather than centralizing it. For regulated industries, privacy-sensitive workflows, and cost-constrained deployments, local inference on open weights unlocks new agentic patterns.