DeepSeek V4.1 Flash aces one-shot Rubik's Cube spatial test; predecessor hit 61% on ARC-2
teortaxesTex · x · 2026-09-15
A user tested DeepSeek V4.1 Flash's spatial reasoning and state tracking: asked to build a Rubik's Cube in Three.js, solve it, scramble it, and continue solving — all one-shot — the model passed with flying colors.
teortaxesTex commented that it has the makings of a model that would do great on ARC-2 puzzles: the previous Flash-0731 already hit 61% on ARC-2 with far more rudimentary non-native multimodality. He notes a strange regression on CritPt and some other benchmarks, and tagged ARC Prize for attention. Combined with the Catan benchmark, community consensus puts the model between GPT-5.6 Terra and Sol at much lower cost.
Related event: DeepSeek V4.1 Flash Matches GPT-5.6 in Catan, Aces Spatial Reasoning Test(2 posts)→
More from Models
- Why OpenAI must copy Google's long game to keep up with Anthropic — haider1 · 2026-11-02
- GDPx: SemiAnalysis's frontier benchmark across the 11 most valuable domains — rohanpaul_ai · 2026-11-02
- SemiAnalysis launches GDPx: best frontier model scores just 12% on expert work — rohanpaul_ai · 2026-11-02
- AWS's open-source Strands Harness claims 45% cheaper than Claude Code, reigniting harness debate — blaizedsouza · 2026-11-02
- Yacine calls Terra High 'kind of an insane model' — yacineMTB · 2026-11-02
- Running DeepSeek V4.1 on a single H200 at $5.30/hr — daluoseo · 2026-11-02