Report: DeepSeek views Kimi's approach as hard to scale, seeks a cheap scalable recipe
teortaxesTex · x · 2026-10-03
Responding to whether DeepSeek and Zhipu hit pretraining walls, teortaxesTex claims DeepSeek considers Kimi's approach worse: the 2.8T model can't be served at sufficient volume at this stage given its compute-heavy architecture, and Zhipu genuinely struggles with scaling. He says DeepSeek wants a recipe that scales while staying cheap. Unverified secondhand claim.
More from Models
- Early tip: set Opus 5.5 to medium — prasenx · 2026-10-03
- Every CEO vibe-checks Fable 5.1: a personal AI feed of every company meeting — every · 2026-10-03
- 11 wild builds with Opus 5.5 and Sonnet 5.5: games, shaders, full videos — socialwithaayan · 2026-10-03
- Unverified: Community-Built Mnemos Memory System Claims 2.5x Accuracy Gain Over Claude Native Memory — RileyRalmuto · 2026-10-03
- Anthropic quietly offers $250 Claude credits bonus and manual weekly usage reset — Orochhe_Marou · 2026-10-03
- Rumor: Deactivated account claims Google has far more powerful internal models than Gemini 4 Argon — mark_k · 2026-10-03