New report details per-dimension forgetting, selective depth retrieval, and a 16-of-896 MoE
Ahmad_Al_Dahle · x · 2026-07-21
A reply highlights three architectural ideas in a new technical report:
- KDA replaces Gated DeltaNet’s single scalar decay with learned per-dimension forgetting.
- AttnRes selectively retrieves across depth instead of weighting every residual layer equally.
- LatentMoE activates only 16 of 896 experts, with routing balanced by quantiles of router scores.
The post frames these as fixes for common one-size-fits-all design choices in linear attention, residual weighting, and expert balancing.
Related event: Moonshot Releases 2.8T Open-Weights Model Kimi K3(14 posts)→
More from Models
- Benchmark scores drop from 89% to 19% on new evals — how benchmaxxing breaks leaderboard trust — airesearch12 · 2026-09-11
- Developer Building a Unified Leaderboard of All Model Benchmark Scores — airesearch12 · 2026-09-11
- Rumor claims Kimi faked performance by serving Claude; DeepSeek new model surprises in evals — realsohamparekh · 2026-09-11
- GPT-5.6 writes well but is instantly forgettable, user complains — BasedRaddka · 2026-09-11
- Opus Refuses Protein Research Codebase Over 'Safety' Concerns, Dev Considers Rolling His Own — josephdviviano · 2026-09-11
- User Hails Unconfirmed 'DeepSeek 4.1 Flash' as an Inflection Point in LLMs — himanshustwts · 2026-09-11