OpenAI ships MedChemBench; Tencent Hunyun team confirms smaller models on roadmap
创业邦 · wechat · 2026-11-22
- OpenAI released MedChemBench, a 99-question benchmark evaluating AI models on medicinal chemistry reasoning: linking molecular structures to experimental evidence, explaining results, comparing compounds, and predicting properties. GPT-6.2Astra scored 84.5% in ultra-high reasoning mode, with GPT-6Astra at 83.3%, GPT-6.1Sol at 82.0%, and GPT-5.6Sol at only 58.2%.
- Nestlé is building toward self-running factories, partnering with Microsoft and Boston Dynamics at a new deep-tech center in Switzerland, with Google talks in final stages and Apple spatial computing tech also in use.
- Tencent Hunyuan Reddit AMA: chief AI scientist Yao Shunyu's team confirmed Hy4-mini and other smaller models are on the roadmap, said a smaller image model is under consideration, and argued Hy3/Hy4 show well-built smaller models can beat larger ones at lower cost.
- Databricks is reportedly in early talks to raise billions at a $225B valuation, up nearly 20% from August's $190B.
- Stealth model Gestalt tops ContraLabs' FrontierDesignBenchmark with Elo 1931 and a 90% win rate, beating ClaudeFable5.2 and GPT-6.1Sol and even human reference works in designer preference — developer unknown.
More from Companies & People
- Obsidian CEO: Top designers wasted prime years on big tech minutiae, AI reawakens their talent — Jackyhuang · 2026-11-22
- Howard Marks: AI's pace of progress makes OpenAI and Anthropic hard to value — WhatTheLJW · 2026-11-22
- OpenAI's first India hire Pragya Misra to lead business for India & South Asia — pragyamisra · 2026-11-22
- Harrison Chase: buying Copilot won't transform teams without a knowledge factory — hwchase17 · 2026-11-22
- Baseten's partner list sparks debate: prompts pass through ~20 GPU middlemen — suchenzang · 2026-11-22
- Analysts clash over Google's TPU split: bet on DeepMind/Gemini or become the Nvidia of TPUs — vitaliychiley · 2026-11-22