Meta has released Muse Glimmer, a 30-billion parameter dense multimodal model optimized for always-on local agent workflows. The weights are available under the Apache 2.0 license.
- Quantized to ~4-bit, the model fits in under 20 GB, leaving headroom for KV cache and speculative decoding on consumer hardware with 24 GB or 32 GB of memory.
- Ships with a lightweight DFlash-based drafter for speculative decoding, enabling significantly faster token generation with identical output quality.
- Trained on 100+ languages with controllable reasoning effort to balance quality and speed.
- Optimized for agentic tasks including end-to-end task completion, function calling, multi-step reasoning, and failure recovery.
The release aims to enable efficient local execution of complex agent loops by reducing memory footprint while maintaining performance on benchmarks like DeepSearch QA and SWE-Bench.