Unverified: Thinky's 2S finished a DeepSeek-Coder-V2-scale pretrain 12x faster and cheaper
teortaxesTex · x · 2026-11-07
AI analyst teortaxesTex cites unposted data to assess newcomer lab Thinky's model progress (unverified).
- Inkling 2S's pretraining run is comparable to DeepSeek-Coder-V2 (June 2024): 21B active params vs 17B, 10.2T vs 12.25T tokens.
- 2S was significantly cheaper (mostly hardware: bf16 vs fp4) and finished about 12x faster.
- He argues 2S is "not remotely in the same league" gap — it's on par with V4.1 if not above, and Thinky's pace makes him reassess industry catch-up timelines.
Related event: Leaks claim Thinky's new model matches DeepSeek at a fraction of the cost(2 posts)→
More from Companies & People
- Microsoft's AI capex hits $45.2B in a quarter while raking in $10B from OpenAI — draginol · 2026-11-07
- Cracked marketers are now more in demand than cracked engineers, says Greg Isenberg — adrianscottcom · 2026-11-07
- AI researcher JinWeiyang announces early-stage cancer diagnosis, surgery scheduled — tianyin_xu · 2026-11-07
- Cofounder retention at AI labs: Anthropic 7/7, OpenAI 3/11, xAI just 1/12 — sahilypatel · 2026-11-07
- FuriosaAI CEO meets Ollama CEO in Seoul to discuss open-weight AI and NPU ecosystem — JungWooHa2 · 2026-11-07
- Viral take: Anthropic employees live 'deeper' lives than OpenAI employees — basedjensen · 2026-11-07