68-hour pretraining: Chinese open-weight models run circles around Western open SoTA
teortaxesTex · x · 2026-11-07
AI commentator teortaxesTex shares an open-weights benchmark table: one model pretrained in just 68 hours plus about a week of post-training delivers stunning results. He argues that for two years random Chinese actors — even "AGI departments" at HR or e-commerce companies — have been beating Western open SoTA, and "Thinky" is another tier above.
Key benchmark highlights (Flash / Pro tiers):
- Kimi K3: GPQA Diamond 93.5%, Apex 2025 65.6%, HMMT Feb 2026 97.0%, AA-LCR v1.1 88.7% — leads reasoning and long context
- GLM-5.3-Flash / GLM 5.3: CrittP v2 57.4%, FORTRESS benign 96.8%, strong on safety
- Inkling 2 Small: best factuality, 73.3% non-hallucination rate on AA-Omniscience
- DeepSeek-V4.1-Flash-1017: top instruction following at 50.4% on ComplexConstraints
- GPT-6.1 reference column still tops CrittP v2 at 65.1%
Related event: Chinese open-source models stun benchmarks, leaving Meta behind(3 posts)→
More from Models
- Only 7 AI products rank top on web, mobile and revenue — Perplexity is one, a16z Top 50 shows — ChrisUniverse · 2026-11-07
- Arena weekly: Kimi K3.1 Pro tops open models, Gemini leads video, GPT Image 2x faster — arena · 2026-11-07
- MiniMax and Fireworks to host SF fireside on building M3.1 Flash for long-running agents — MiniMax_AI · 2026-11-07
- Pre-training ladder trick: test every recipe change on 300M–6B models first — teortaxesTex · 2026-11-07
- 'Hidden GPT-6.2 Astra' claim debunked: fake reveal tricked via one-line console object — teortaxesTex · 2026-11-07
- Inkling 2 Small tops open-weight models on ARC-AGI-2 at $0.27/task, weights coming — LiTianleli · 2026-11-07