UniPat AI
Evaluation-first domestic expert-data company that also trains its own science and math models
Overview
- Founded: 2025
- HQ: Beijing
- Funding: Series A of $300M reported in 2026-09, led by Alibaba, with Tencent and HongShan China (or Hongtai, per some reports) participating, valuation ~$2.5B; Bloomberg says the deal is still under negotiation and terms may change (media reports, ⚠️ no official announcement seen)
- Links: https://unipat.ai
This profile was compiled from the AI training-data column cards (LLM); each column's data view is shown below under "Data profile" (in Chinese).
UniPat AI
评测优先的国内专家数据公司,同时自训科研与数学模型
- 数据类型
- 评测基准;合成数据;专家 rubric;RL 环境
- 规模口径
- UniScientist 数据集 4,700+ 研究级样本、50+ 学科、每样本专家 1–2 小时(公司口径)
- 客户
- 未披露
做什么
UniPat 由前通义团队成员李宽创办,不只是评测公司:自建 Monthly-SWEBench、SaaS-Bench(23 个 SaaS 系统、106 任务)[5]、RoadmapBench、EvoCode-Bench 等基准,另训 UniScientist-30B-A3B 与 UniMath-35B-A3B,二者均建在 Qwen 权重上 [2][3]。基准列表见 vendor-benchmarks。
数据怎么产
Evolving Polymathic Synthesis:LLM 当跨学科出题者,人类专家当高精度验证者,每题 20+ 条可验证 rubric,开放式报告被拆成多个封闭式检查(公司口径,无第三方复测)[2],思路见 rubric-rewards。UniMath 自报 IMO-ProofBench 从基座 23.1% 到 86.0%,但两列评分口径不同。
采用度与利益冲突
截至 2026-09-20,9 个公开仓库中仅 4 个带明确许可证 [4],未见头部模型技术报告引用其基准;UniScientist 的 HF 下载量仅 75 次。基座、资金与团队均与阿里相关,而榜单为 qwen 及竞品打分,偏差不需作弊即可产生。“与 Qwen 采购关系”无证据。同梯队见 humanlaya。