Project Numina
French non-profit building open AI-for-math — NuminaMath dataset, Kimina-Prover, and Numina-Lean-Agent for formal theorem proving.
1. Core Product / Service
Project Numina is a French non-profit and open scientific collaboration dedicated to advancing mathematical AI, releasing all research publicly under a public-benefit mandate [1]. Its output splits into three layers:
- Datasets — NuminaMath, the largest public corpus of mathematical problems and proofs (~860,000 entries), plus NuminaMath Lean 100k and CombiBench; the roadmap targets formalizing up to 1 million problems/proofs [1].
- Models — NuminaMath 7B, which won the AIMO Progress Prize #1 in July 2024, and the Kimina-Prover line, developed jointly with Moonshot AI's kimi team [1].
- Agents — Numina-Lean-Agent, an open agentic reasoning system for formal mathematics (Lean), which solved all 12 problems of the 2025 Putnam competition and formalized the Brascamp–Lieb theorem [1][3].
Numina's strategic fork can be framed as self-trained proof models vs. public-compute infrastructure — an open, grant-funded path rather than a closed lab or VC bet [1].
2. Target Users & Pain Points
- Research mathematicians and formal-methods community needing open, reproducible theorem-proving tooling.
- Olympiad / competitive-math educators — NuminaMath and the fellowship explicitly target educational accessibility.
Pain solved: formal proof is expensive to produce and verify; Numina's open datasets and models lower the cost of entry and avoid locking math verification behind a proprietary API.
3. Competitive Landscape
| Player | Type | Model | Notes |
|---|---|---|---|
| Numina | Open non-profit | NuminaMath, Kimina-Prover | Open datasets + agents; AIMO Prize #1 |
| google-deepmind | Closed lab | AlphaProof / AlphaGeometry | IMO silver (2024) |
| axiom | VC-backed startup | AxiomProver | Putnam 12/12, $264M raised |
| Kimi (Moonshot) | Chinese lab | Kimi-Prover | Numina collaboration partner |
| ByteDance Seed | Chinese lab | Kakeya | Numina collaboration partner [1] |
Numina's differentiation is openness and non-profit governance versus DeepMind's closed frontier and axiom's commercial "Verified AI" push.
4. Unique Observations
- The "Kimina" name encodes the alliance. Kimina-Prover is the Numina × Moonshot (kimi) joint prover; the collaboration timeline links Numina with both Kimi (Kimi-Prover) and ByteDance Seed (Kakeya), an unusual cross-lab, cross-border open-math coalition [1].
- Grant economics over VC economics. Funded by an €3M XTX Markets research grant rather than venture rounds, Numina is structurally insulated from the benchmark-hype incentives that drive axiom-style startups — a genuinely different answer to "who funds formal math AI" [1].
- Formal verification as public good. By open-sourcing datasets and agents, Numina bets that math verification is infrastructure, not a moat — the opposite thesis of closed "Verified AI" startups [1][3].
5. Financials / Funding
- €3M research grant from XTX Markets (announced Dec 17, 2024) — to formalize up to 100k+ items and open-source reasoning models [1].
- $131,072 AIMO Progress Prize #1 (July 2024) with NuminaMath 7B [1].
- Numina Fellowship — 12–18 month research collaborations with $50k–$200k per project in compute + tooling [2].
- No disclosed venture rounds or valuation (non-profit).
6. People & Relationships
- Co-founder / President: Yann Fleureau.
- Co-founder / Chief Scientist: Jia Li (Olympiad coach; curated validation sets and training data).
- Founding partner: Alexandre Momeni (ex-Stanford ML research scientist, ex-General Catalyst) [4].
- Early team / advisors: Hélène Evain; Guillaume Lample (Mistral co-founder) and Stanislas Polu reportedly aided the founding.
- Partners: Moonshot AI (kimi, Kimina-Prover), ByteDance Seed (Kakeya) [1]; grant funder XTX Markets.
- Competitors: google-deepmind, axiom.
Sources
[1] Internal research notes — Numina core product, self-training vs public-compute route, Kimi-Prover + ByteDance Seed (Kakeya) cooperation and timeline, team. [2] Project Numina, "Numina Fellowship," https://projectnumina.ai/fellowship (2026-10-05) [3] "Numina-Lean-Agent: An Open and General Agentic Reasoning System for Formal Mathematics," arXiv 2601.14027, https://arxiv.org/pdf/2601.14027v1 (2026-10-05) [4] Crunchbase, "Alexandre Momeni — Founding Partner @ Project Numina," https://www.crunchbase.com/person/alexandre-momeni-20c4 (2026-10-05)
Project Numina(NuminaMath / Kimina-Prover)
法国非营利的开放数学数据与证明器,资产是数据加验证层
- 数据类型
- 数学题与解答;形式化 Lean 证明;推理轨迹
- 规模口径
- NuminaMath-CoT 约 86 万条、NuminaMath-1.5 约 90 万题;NuminaMath-LEAN 约 10.4 万条(项目口径)
- License
- Apache-2.0(NuminaMath-CoT 与 NuminaMath-1.5 数据卡标注)
内容与规模
Project Numina 是法国的非营利开放协作项目,在 AIMO 进步奖中以 NuminaMath 7B 夺得第一届冠军(2024-07)。NuminaMath-CoT 发布于 2024-07,NuminaMath-1.5 于 2025-02,约 90 万道竞赛级数学题(CoT 版约 86 万条),Apache-2.0 [1][2]。另有 NuminaMath-LEAN(形式化数据)。
与 Moonshot 合作
Kimina-Prover 由 Numina 与 Moonshot(Kimi)团队合作,基于 Qwen2.5-72B 与大规模 RL;配套的 Kimina Lean Server 在 RL 中充当奖励验证层。
谁在用
NuminaMath-TIR 的 64,312 条进入 tulu-3 SFT 混合;面壁等中国团队的 RL 数据也引用 Numina 题集;Skywork-OR1 的 RL 数据源自 NuminaMath-1.5 等公开集 ⚠️ 为第三方综述口径。
口径提示
项目由 XTX Markets €300 万资助 [3];它不训练面向终端的大模型,价值在数据与验证层,见 rlvr。