Fields Model Initiative

Background

The story

The AIMO Proof Pilot was the first competition dedicated solely to fully open-source LLMs that solve Olympiad-level math problems, and the models that came out of it rival Opus 4.8. After the competition closed, Simon Frieder of the AIMO Prize brought together the three winning teams and researchers from the National Institute of Informatics (NII) in Japan to develop the winning model further, under the umbrella of the Fields Model Organization. The improved model was published before IMO 2026 began, ruling out contamination, and then graded on the competition problems by IMO medallists — the cleanest evaluation of its kind to date. The harness and the full code are public.

FM-Pochi-32B is the joint effort of the following individual contributors: Yi-Chia Chen, Chan Kha Vu, Geremie Yeo, Manh Nguyen, Jacob Morrison, Sam Bealing, Michal Štefánik, Adam Vawda, Pontus Stenetorp, Liu Chaoran, and Simon Frieder.

This effort is the first time that a fully open-source LLM achieves a medal at the International Mathematical Olympiad 2026 (IMO 2026). This has for a long time been a cornerstone for mathematical evaluation, and the series of AI Math Olympiads (AIMO) have as their objective to support the creation of these models. The current result in fact builds upon the latest AIMO competition, the AIMO Proof Pilot.

Significance

Why this is a big deal

For the first time it is possible to do detailed post-mortems. We hope that this will inspire other people to build more on our work and that this kickstarts new data attribution efforts and introspection efforts.

While the first landmark result was the AIMO2 × OpenAI evaluation in March 2025, a flood of further news followed with various claims about LLMs being able to solve IMO-level problems. Most of these are very hard to verify. In particular, a lot of nuance is not shared: How expensive is it to carry these evaluations out? What hardware does one need?

From the point of view of science and scientific reproducibility, most of these were less rigorous than the AIMO2 × OpenAI evaluation, which had more unseen, difficult Olympiad-level problems than the IMO (50 instead of 6), and a controlled time span of several hours. Due to the large number of evaluated problems, it would have been very hard to, e.g., adjust some settings on the fly during the evaluation to improve performance. The AIMO2 × OpenAI evaluation was therefore the first point in time that members outside of AI companies could see that LLMs were able to solve unseen math problems.

Our contribution

What we did

We shared everything: Not just base models. Not just the outputs. But also experiment logs and information about the hardware. Everything is fully reproducible. Our effort is the first time that results that have been claimed for over a year are fully made accessible.

Evaluation Transparency Audit · IMO 2026

Reproducibility Comparison

#	Model / System	Score	Medal	Open weights before eval?	Exact checkpoint public before eval?	Proofs shared?	Reproducibility
1	Huawei Celia	42/42	Gold	No	N/A	No	E
2	RedNote / Xiaohongshu dots-note-3.0	42/42	Gold	No	N/A	Partial (PDFs)	E
3	NVIDIA Nemotron 3 Ultra ensemble	30/42	Gold	Partial	No	No	E
4	Meta internal Muse Spark	Gold (top 4%)	Gold	No	N/A	No	E
5	Claude Opus 5 — Anthropic eval	42/42	Gold	No	N/A	Partial	C
6	Claude Fable 5 — Deedy eval	42/42	Gold	No	N/A	Yes	C
7	GPT-5.6 Sol xhigh — Deedy eval	39/42 (42/42 after repair)	Gold	No	N/A	Yes	C
8	GPT-5.6 Sol Pro xhigh — Deedy eval	37/42	Gold	No	N/A	Yes	C
9	GPT-5.6 Sol max — Deedy eval	30/42	Gold	No	N/A	Yes	C
10	GPT-5.6 Sol default — Deedy eval	28/42	Silver	No	N/A	Yes	C
11	Kimi K3 — Deedy eval	36/42 (42/42 after repair)	Gold	No	N/A	Yes	C
12	Meta Muse Spark 1.1 — Deedy eval	26/42	Silver	No	N/A	Yes	D
13	DeepSeek V4 Pro — Deedy eval	19/42	Bronze	Yes	No	Yes	C/D
14	Claude Sonnet 5 + AutoFyn	35/42	Gold	No	N/A	No	C
15	GLM-5.2 + AutoFyn	34 / 34 / 28 (three runs)	Gold / Silver	Yes	No	Partial	C
16	FM-Pochi-32B (ours)	21/42	Bronze	Yes — fully open	Yes	Yes	A

Reproducibility grades: A = fully self-contained and reproducible; C = protocol documented, manual reconstruction required; D = outputs auditable, generation not plug-and-play; E = effectively non-reproducible.