100 个编程环境实测：M3.1 Flash 价格接近 MiMo V2.6 Flash 但性能更低

Leo Linsky @leo_linsky原推发布于 2026.10.27

在 X 上打开原推

We evaluated Kimi K3.1 Pro and Minimax 3.1 Flash across 100 coding environments. Overall Kimi K3.1 Pro is not an upgrade in overall intelligence (better in fluid intelligence, worse than K3 in long horizon agentic coding) -- within margin of error of the previous release, but it achieved that similar performance with far fewer reasoning tokens at the same effort level, which combined with the new pricing leading to a 60%+ price reduction! However, MiMo V2.6 Pro is still the most cost competitive open weights model at a comparable intelligence today. This is a promising direction for Kimi, as they basically approached the reasoning length problem like Ember-1/Fireworks, but they did so without much of a performance hit (performance is roughly neutral/within margin of error of Kimi K3 at a significant price/speed boost). For Minimax M3.1 Flash, it ends up pricing similarly to MiMo V2.6 Flash, but at lower performance. It would have been a strong release 3 months ago, but today it's nowhere near the Pareto front for open weight or proprietary models.