Kimi bets RL on a 3T base: trajectory volume beats model capacity, DeepSeek disagrees
teortaxesTex · x · 2026-10-23
teortaxesTex lays out a scaling hypothesis: an extremely inference-expensive base model may be the fastest route to RSI via large per-step gains, while DeepSeek's stance is that at iso-flop/iso-bit, returns to trajectory volume currently exceed returns to model capacity.
Screenshots add context: Wu Haoning confirms "RL on a 3T model is amazing"; Artificial Analysis scores Kimi K3 (2.8T params) at 57 on its Intelligence Index, comparable to Opus 4.8 and GPT-5.5, with Moonshot planning open weights; DeepSeek-V4 paper (V4-Pro 1.6T/49B active, V4-Flash 284B/13B active) achieves 1M-token context via hybrid sparse attention.
The author argues Kimi can't afford to be wrong here.
More from Models
- Recursive's Flint: Apache 2.0 local coding model hits 125+ tokens/sec on a MacBook Pro — ChengleiSi · 2026-10-23
- Neolab ecosystem ships flurry of updates: $1B Series A, new coding model Flint — FinanceYF5 · 2026-10-23
- Recursive Flint benchmarks: matches Qwen3.8-27B on agentic tasks, weak on long context — ChengleiSi · 2026-10-23
- Kimi K3.1 Rumored to Launch Within ~24 Hours — ChrisUniverse · 2026-10-23
- Redditor open-sources behavioral architecture of his Monday custom GPT with 30% VRAM-saving trick — Top-Koala5617 · 2026-10-23
- Goshen Labs teases Ras Class models built on Qwen3.8-27B, coding-focused variants — Rasmic · 2026-10-23