Teortaxes: Kimi stuck serving old K2 base as K2.8 rumored post-trained variant; DeepSeek's compute moat
teortaxesTex · x · 2026-09-11
Teortaxes argues that if Kimi K2.8 is really a post-trained K2.7 (ultimately the July 2025 K2), it confirms Moonshot struggles to pretrain new base models and must keep serving the relatively expensive old K2 base — while DeepSeek's ownership of compute lets it ship V4 and then V4.1. A replier questions the economics of sticking with the costly K2 base absent K3, speculating a sparse-attention workaround may be in the works.
More from Models
- DeepSeek V4.1 Flash reportedly bakes prefill/decode disaggregation into the model weights — altryne · 2026-09-11
- Gemini Spark is now available in Italy — Individual-Sorbet489 · 2026-09-11
- Qwen core dev teases 'v4p back? k3-0.2?' in cryptic model hint — JustinLin610 · 2026-09-11
- Same Codex task, measured: GPT-6 Astra Low costs 6.9x more than Sol High — arty0mk · 2026-09-11
- Frontier model weights are near-impossible to steal or self-replicate, argues Bindu Reddy vs AI doomers — bindureddy · 2026-09-11
- Nathan Lambert publishes definitive open-models reading list: gap now just 4-6 months — Interconnects (Nathan Lambert) · 2026-09-11