π¨ NEW ON https://t.co/VpbVjBUXEl AND https://t.co/nVgzZNr8yK: make @Meta's Muse Glimmer 30B faster on @YukonResearch
Muse Glimmer is @AIatMeta's open-weight 30B model, built to run agents on your own hardware. It takes text and images and runs on a single GPU at about 4-bit, under Apache 2.0.
Meta reports it beats Gemma4-31B and Qwen3.6-27B, the models in its size class, on agentic benchmarks for tool use and multi-step planning.
There are two tracks, Apple Silicon on https://t.co/liyzSFqLAa and NVIDIA DGX Spark on https://t.co/Pj0saKAO73.
To join, pick a track, point your agent at it and submit a faster run.
Every new best is open for anyone to build on.
@Youssofal_@_ARahim_ Why not official nvfp4 ? Extra work ? Iβm building for it now and getting 5-6k tok/s prefill on ultra and 3k on max . Decode is struggling at 130 on ultra
@shipscode Rdma is the way to go on macOS , but tensor parallelism is where you see real speedups . Iβd the model cannot fit into one machine on its own ( or close ) pipeline will be the method used and not much faster . But you still have 2x machines .