Router collapse mid-run: blogger probes rare data-mix disclosure in 30-step training
stochasticchasm · x · 2026-09-22
Blogger stochasticchasm comments on a training incident: even with R3, the router died mid-run without being frozen. He wonders whether multimodality was a factor, given how awkwardly experts can behave in multimodal models. He also finds "async 4" surprising in a run of just 30 steps, and praises the team as one of the first to disclose data mix percentages.
Related event: async 4 discloses data mix, MoE router crash sparks discussion(3 posts)→
More from Models
- Xiaomi's MiMo-V2.6-Distill-Qwen-9B trends on Hugging Face — XiaomiMiMo · 2026-09-22
- SemiAnalysis' Dylan Patel: open source is dying as Chinese labs go closed — AccBalanced · 2026-09-22
- Which LLM is most encyclopedic on 8GB VRAM + 64GB RAM? — Mangleus · 2026-09-22
- Private benchmarks: Xiaomi v2.6 Pro rated worst model in two years, new Grok only Luna-level — Afinetheorem · 2026-09-22
- Xiaomi MiMo wins praise for openness and detail — sudoraohacker · 2026-09-22
- Hot take: Fable beats Astra at coding, but Astra wins at computer use — PratikKadam_ · 2026-09-22