Kimi K3.1 Pro becomes first open-weights model to score on MazeBench, at just 2%
patience_cave · x · 2026-10-25
Blogger patiencecave tested Kimi K3.1 Pro against DeepSeek v4.1 on MazeBench, a multistep maze benchmark. Kimi K3.1 Pro solved a multistep puzzle on its 20th attempt — only 5 other models had solved it before — becoming the first open-weights model with a nontrivial score of 2%, with DeepSeek v4.1 Flash close behind at 1%. The near-zero scores underline how hard long-horizon planning remains for all models.
Related event: Kimi K3.1 Pro Becomes First Open-Weight Model to Score on MazeBench(3 posts)→
More from Models
- Yacine torches 'open weights': it's a fake term that just means closed source — yacineMTB · 2026-10-25
- What can a 4GB model do on a robot? Nemotron 3.5 Bolt E4B tested on Jetson — lmoroney · 2026-10-25
- NVIDIA un-nerfs consumer cards: 'you bought the card and get the whole card' — TheZachMueller · 2026-10-25
- MLX dev's top 3 local coding models: Flint, Cohere and Qwen 27B — thursdai_pod · 2026-10-25
- MazeBench Chart: Open-Weight Models Still Far Behind GPT-6 Astra (16% vs 1-2%) — patience_cave · 2026-10-25
- DeepSeek v4.1 Flash Makes Big MazeBench Leap Over Predecessor, Near Kimi's 3T Model — patience_cave · 2026-10-25