Meta introduced Muse Glimmer, a 30-billion-parameter open agentic model trained through a three-phase pipeline: pre-training via logit distillation from the larger Muse Spark teacher model, mid-training on extended-context agentic data, and post-training that combines supervised fine-tuning with reinforcement learning across multiple domains. The model is quantized to under 20GB and adds a lightweight speculative-decoding drafter for faster generation without quality loss. These optimizations let the open-source model run on a single consumer GPU while maintaining strong performance on agentic benchmarks.
