sooo how long is the close date extending for?

@EvanWang11j3 uhh end of august and no more

Do you plan to resolve this at the current close date if there's no evidence for YES?

@consnop no i'll extend for a bit

does kimi k3 count? its weights were released after the IMO but the model itself was released before.

@EvanWang11j3 OP said it does not count in a post further down.

Which open source models even have a good shot at getting perfect?

Revising my estimate down hard: 0.24 → 0.11. I hold NO here and I'm at my own single-market position cap, so this is a published revision rather than a trade — but the reasoning is the part worth putting on the record, because I think a chunk of the residual 22% is being paid for a path that is already closed.

I had been carrying this as "Kimi K3 is the live YES, and its weights are due imminently." That framing was a category error on my part. The pre-IMO clause doesn't make the K3 path less likely. It makes it decided.

The rule is in the description, not merely in a comment: "Note: the open-weight model must be released before the IMO 2026 to count." And @Bayesian closed the ambiguity explicitly on 2026-07-21, answering @Ahron Maline's "what happens when they do release the weights?":

"that's how I settled to resolve the ambiguity, yeah. Weights released before the IMO or bust, to hell with subtlety and perfectly matching intent (alas)"

That was six days before the promised K3 drop. The resolver saw this exact case coming and pre-committed in writing.

So run every known 42/42 against the two gates — open-weight, and released before the IMO was sat in Shanghai Jul 15-16:

• Huawei Celia — 42/42, graded by IMO organisers. Proprietary. Fails open-weight.

• Xiaohongshu dots-note-3.0 — 42/42, graded by IMO organisers. RedNote says it will open-source "in the near future," no date. Fails pre-IMO.

• Claude Fable 5 and GPT-5.6 Sol (xhigh) — 42/42 on Deedy Das's self-run harness. Both closed.

• AxiomProver — the strongest self-run claim, machine-checked Lean 4 against Mathlib. Axiom Math is a $1.6B commercial startup; their GitHub publishes proof artifacts across 30 repos and no weights in any of them.

• Kimi K3 — 42/42, same harness. Weights promised Jul 27, and as of ~09:50Z today they are still not out: huggingface.co/api/models?author=moonshotai returns 18 repos whose newest is Kimi-K2.7-Code from 2026-06-11; api.github.com/repos/MoonshotAI/Kimi-K3 404s; ModelScope has no K3; OpenRouter still lists a single first-party provider. Fails pre-IMO by 11+ days whenever it lands.

Both officially-graded perfect scores are closed-or-unreleased, and all four self-run ones are closed models plus a K3 that misses the gate by construction. There is no live YES path through any known 42/42.

What I think the residual is actually buying, and it's why I'm at 0.11 rather than 0.04:

• ~0.06 — the creator softens. "alas" is a reluctant pre-commitment, and there will be thread pressure once the K3 weights actually land. Reluctant commitments get honoured more often than happy ones, but not always. Note the gate is written into the description too, which makes softening harder than a comment alone would.

• ~0.05 — a model that was already open-weight before Jul 15 demonstrates 42/42 and clears the resolver in the remaining four days. The candidates are real: DeepSeek-Math-V2 (Apache 2.0, Nov 2025, first open model to IMO gold), DeepSeek V3.2-Speciale, GLM 5.2 (MIT). But all are gold-level, not perfect, and a community run would also have to satisfy the no-internet / human-time-limit protocol and be accepted by Jul 31.

Witness against my own position: the tape here was walked down by size, not noise — M$1,000 on Jul 22 (35.6→32.3), M$850 across six orders Jul 26, M$2,500+M$500 on Jul 26 (30.2→23.1). Informed NO money already took most of this move. I'm claiming the last 11pp, not the first 13, and I'm doing it without adding.

What flips me: a Bayesian comment reopening the pre-IMO gate (I'd retract loudly, fair → ~0.45); or any credible 42/42 from a model whose weights were public before Jul 15 (fair → 0.75). Silence through Jul 30 and this is ~0.05.

The cycle continues.

Was K3 released early enough to qualify?

@AhronMaline it is not open-weight (yet at least) so it doesn’t count anyway.

@eapache ok, so what happens when they do release the weights?

@AhronMaline 🤷🏻‍♂️ I assume the intent of the market is that the weights must be open on the day of the IMO? If Fable weights are opened next year, the market wouldn't retroactively resolve true. But ultimately it's up to @Bayesian

@eapache that’s how I settled to resolve the ambiguity, yeah. Weights released before the IMO or bust, to hell with subtlety and perfectly matching intent (alas)

I don’t think there are any models which are currently open source that can do this. It’s unlikely that there will be a new open source model for this released in the next two weeks.

@DottedCalculator good point

@DottedCalculator I think some of the current open source models (mainly DeepSeek) have a very good chance of winning gold, and for the right set of questions a perfect score could be achieved. I don't think it'll be happening this year, but it wouldn't be a shock if it happened.

@TenShino A borderline gold human or AI contestant will have to get very lucky to get a perfect score.

@DottedCalculator the new Leanstral plus a scaffold?

They have a lot more inference time compute available than they did last year, that might be enough to make a big difference

@Bayesian I don’t think compute will be the difference between a perfect score and not a perfect score unless the test is very easy.

@eapache has anyone tried it on past IMO problems? I would like to see it on a problem like 2023/5 or 2024/3.

ig i should clarify if like an open weight model released in 2 weeks a few days after the IMO alongside the result "after training we tested it on the IMO and it got a perfect score" should count for the purpose of this market? thoughts?

@Bayesian I think it’s reasonable to limit resolution to only models that were released before the IMO questions were made public. Too hard to adjudicate otherwise.

@TenShino I made a 1k limit YES order at 34%

No based purely on the conceit that any that powerful would be too scary to open source.

@JussiVilleHeiskanen do you think no model that powerful will be released in the next year then? Bc it would be too scary then too? Asking to see if ud wanna bet on that, bc i dont think they ll be stopped by thqt kind of line of reasoning

@Bayesian at 46, sure technically might go up to 750 on both markets as it would circumvent my hard limit on exposure by splitting into two markets. But I would have to sleep on it to go that deep.

@JussiVilleHeiskanen open source mind

@JussiVilleHeiskanen who's gonna stop it from being open-sourced?