Moonshot’s Kimi K3 report says sparse MoE and delta attention beat bigger dense models
SumitGup · x · 2026-07-26
Moonshot AI’s Kimi K3 technical report argues that bigger dense models are not the future, and instead highlights five design ideas: sparse routing, delta attention, attention residuals, a 1M-token context window, and efficient scaling.
- Sparse MoE routing: 896 experts exist, but only 16 are activated per token.
- Delta attention: focuses computation on what changed instead of recomputing full context.
- Attention residuals: preserve useful signal across layers.
- 1M context: aims to fit entire repositories, books, and papers in one session.
- Efficient scaling: targets frontier coding performance without frontier inference costs.
Related event: Moonshot AI Unveils Trillion-Parameter Kimi K3(4 posts)→
More from Models
- 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
- Terminal Bench v4: GLM-5.3 Leads at 41.9%, Kimi-K3 Underwhelms at 12.6% — Ok_Warning2146 · 2026-09-11
- GPT-6 Astra beats Factorio with enemies in 44 in-game hours at ~$4,500 API cost — liminal_bardo · 2026-09-11