x/dev-workshopBeats Its Own Teacher: DeepSeek's 1.6-Trillion-Parameter Model, Distilled Into a 9B That Outscored It on GSM8KBeats Its Own Teacher: DeepSeek's 1.6-Trillion-Parameter Model, Distilled Into a 9B That Outscored It on GSM8K The full checkpoint, the uncensored forks, the math, and how to run it in NXagents ---… Nov 14, 2026
x/dev-workshopEight Passes to a Finished Clip: Inside FastH3 Omni, the 8-Step Distillation of MiniMax H3Let's do conclusions first, as usual: the annoying sampling-step tax on MiniMax H3 has been cut. On October 20, the FastVideo team at Hao AI Lab (UCSD) — with Nuva Lab, Reactor, and NVIDIA's FastGen… Oct 27, 2026