Kimi's base-model struggles vs DeepSeek's compute ownership advantage
teortaxesTex · x · 2026-09-11
Analyst teortaxesTex argues Moonshot is doing RL on K3 but re-post-training an old MLA base (K2.7, ultimately K2 from July 2025), suggesting real scarcity issues and over-dependence on Alibaba's compute. In contrast, DeepSeek's ownership of compute lets it ship V4s and V4.1 quickly. Key thesis: owning your compute matters for base-model iteration speed.
More from Companies & People
- Yann LeCun shares World Models update at ECCV 2026 — AntonObukhov1 · 2026-09-11
- Thursdai podcast roundup: Cognition SWE-2, Lyria 2.5, CoreWeave agent hackathon — thursdai_pod · 2026-09-11
- ThursdAI: Cognition SWE-2 nears frontier, safety week drama and OpenAI's Navier-Stokes claim — thursdai_pod · 2026-09-11
- Tennis creator with 1M followers uses Lovable to run global community events — damienghader · 2026-09-11
- NSF launches new postdoc fellowship for AI x Bio; PI offers to sponsor applicants — pastramimachine · 2026-09-11
- ThursdAI Sep 10: DeepSeek V4.1 Flash, OpenAI's Millennium claim, Meta Muse, rough safety week — thursdai_pod · 2026-09-11