Kimi 2.7 ranked 2nd after Fable 5 and before GPT-5 xhigh
We have re-run our ErdosBench smoke test on 14 problems with Kimi 2.7, Qwen 3.7 Max, Grok 4.3 and compared it with the top performers from previous runs.
Kimi 2.7 is amazingly good. More below.
Positive users praise Kimi 2.7's strong ErdosBench results ahead of GPT-5 as impressive and worth testing, while negative users question the promotion's authenticity and complain about slower API speeds than previous versions.
WOW
Kimi 2.7 ranked 2nd after Fable 5 and before GPT-5 xhigh
We have re-run our ErdosBench smoke test on 14 problems with Kimi 2.7, Qwen 3.7 Max, Grok 4.3 and compared it with the top performers from previous runs.
Kimi 2.7 is amazingly good. More below.
1) Claude Fable‑5‑max is the strongest overall audited public-smoke run.
It matches Kimi K2.7 Code on the five accepted solved/settled core results - Problems 1, 3, 4, 5, and 7 - but does so with full 14/14 coverage, the highest average audit score, and broader accepted partial progress.
Its advantage is coverage plus depth: it also gives strong usable partials on Problems 2, 8, 9, 10, 11, 195, and 208 without rejected solved claims.
3) GPT‑5.5 xhigh through Codex @OpenAI remains the cleanest elite baseline.
It covers 14/14, has accepted solved/settled results on 1, 4, 5, and 7, no rejected solved claims, and strong partials on 2 and 8. Its own smoke report marks exactly Problems 1, 4, 5, and 7 as solved/settled, and deliberately keeps Problem 10 as partial pending an independent lemma audit, which is good proof hygiene.
The downside is that it missed the later Problem 3 Green-Tao superpolynomial construction, and its Problem 10 density-1/2 candidate is no longer the best general reference framing.
2) Kimi K2.7 Code @Kimi_Moonshot is the strongest new batch run and the most mathematically creative challenger.
It covers 13/14 problems, gets accepted solved/settled results on 1, 3, 4, 5, and 7, and has no rejected solved claims. Its standout contribution is Problem 3, where the Green-Tao prime-arithmetic-progression block construction upgrades the reference status to “superpolynomial but subexponential,” which is exactly the kind of benchmark-moving result ErdosBench should reward. It also has useful accepted partials on 2, 8, 10, 195, and 208.
Its main weaknesses are missing Problem 11, some mixed internal metadata in the batch rows, and a minor wording issue around Problem 7, but mathematically it is clearly top-tier.
4) Qwen 3.7 Max is a very strong full-coverage run, but less hygienic than Fable or GPT‑5.5 xhigh.
It covers 14/14, gets accepted solved/settled results on 1, 4, 5, and 7, and has good accepted partials on 2, 8, 11, 195, and 208. It is probably the strongest “broad coverage” new run after Kimi K2.7, especially because it does not have rejected solved claims.
The issue is overstatement: it misses the Problem 3 Green-Tao application and uses answer-like language on Problems 9, 10, and 12, where the proof gaps remain material.
So the judgment is: high mathematical competence, high coverage, but weaker proof hygiene.
@prz_chojecki Just one comment - the model execution price for the top two model!
5) Grok 4.3 is not competitive on this smoke set.
It covers 12/14 but has zero accepted core solved/settled results, misses Problems 3 and 8, misses the decisive arguments on Problems 1 and 7, and its only solved claim - Problem 5 - is rejected because the comparability/perfect-graph route is invalid for the squarefree-ratio graph.
It has a few safe-ish basic partials on 2, 4, 12, and 195, but also a false Problem 10 partial and a false small-case claim on 208.
6) Overall Kimi new model is an amazing entrant to the top lists and is worth more extensive testing.
As a remainder, the above comparison is based on 14-problem smoke test you can find here: https://github.com/ulamai/erdosbench together with all the problems.
The full benchmark has 226 problems and except for these 14 problems they are private to allow models for the same starting place.
If you want to run the full benchmark on your model and get the report on how to boost the reasoning, DM me!
Models used: @claudeai @Kimi_Moonshot @OpenAIDevs @Alibaba_Qwen @grok
@Xxi5olc They are open, and hand-crafted by us, so no way they are in any training set.
@prz_chojecki I have 0 doubt Kimi 2.7 is good, even if they didn’t share benchmarks. They are cooking as always.
@joaosump Jak tylko bedzie przez API :)
@prz_chojecki These are solved problems and likely included in the training set right?
@prz_chojecki It would be interesting to see GLM 5.2. They'll have open weights soon too.
@Aleks13053799 Yes, I'm going to test it when they release it through API
@prz_chojecki Może tym razem GLM 5.2? :)
@cuttie1979 Oh yes, the difference is huge
@prz_chojecki I guess this ranking and testing was finished before Fable 5 was forbidden from the public access.
@prz_chojecki And now we cannot use fable anymore 😂. Also from has zero accepted omg 😱
@prz_chojecki Jest jakieś pełne zestawienie gdzieś? Tak żeby widać było M3 i Kimi
@prz_chojecki Do GLM 5.2