Kimi K3.1 reportedly launching mid-October with long-horizon agent focus; K2.8 pricing raises concerns
teortaxesTex · x · 2026-10-10
- Kimi K3.1 from Moonshot is reportedly set for a mid-to-early-October launch, focused on long-horizon agent capabilities.
- Training uses computer-use logs plus real task data to fix dropped long task chains; a 5-6T parameter model is also in progress, bottlenecked by high-end training compute (domestic chips fine for inference, not yet for pretraining).
- Strategy pairs a flagship model with lightweight K2.8 for coding and daily tasks.
- teortaxesTex argues that if K2.8 is another K2+, its inference cost already exceeds any DeepSeek model since April — worrying Moonshot may have scaled too early.
More from Companies & People
- Three mistakes startups make hiring head-of-creator roles, from someone repeatedly offered them — FinanceYF5 · 2026-10-10
- Stanford team lists published work: triage, labs, discharge summaries — jonc101x · 2026-10-10
- Ex-engineer reveals offline Siri he built 8 years ago beat the cloud 25% of the time — jdluk87 · 2026-10-10
- George Hotz publishes new blog post 'Lobbying' — smilelamp · 2026-10-10
- Bezos prefers 'superintelligence' to AI: 'artificial' sounds fake — rohanpaul_ai · 2026-10-10
- YouMind founder Jade on podcast: how AI apps survive being swallowed by foundation models — sven_ai · 2026-10-10