DeepSeek V4.1-Flash rebuilds an Adidas shoe in Blender — and honestly lists 9 mismatches vs the reference
cedric_chee · x · 2026-10-18
A head-to-head eval pitting DeepSeek V4.1-Flash-1017 against Opus 5.5, GPT-6.1 Sol, Astra, and Fable 5.1 on one task: recreate the Adidas Adizero Adios Pro in Blender from studio shots, then build a real brand website — comparing cost, efficiency, 3D accuracy, render fidelity, and design taste.
The screenshot's highlight is the model's BUILDNOTES: it admits "the construction logic and page are much stronger than the sculpt" and candidly lists 9 things it could not match — toe box volume, heel/collar sculpture, tongue, stripe geometry, upper translucency, foam surface, outsole, ENERGYRIM construction, and the missing true cutaway — each with reasoning (e.g., transparency washed out black stripes to gray, so it settled at 0.96 alpha).
QA evidence: no console errors, 26 draw calls, 335k triangles, 2048 shadow map, 99.7% cache hit, 54.6M tokens used. Author's verdict: "at least DeepSeek is honest about the gap."
Related event: DeepSeek beats Opus in blind Blender benchmark at fraction of cost(3 posts)→
More from Models
- Claude Opus simulates a user explaining Opus to itself, then breaks in with an 'inner-er voice' — repligate · 2026-10-18
- DeepSeek V4.1-Flash-1017 outscores Claude Opus 5 on three agentic coding benchmarks — eyishazyer · 2026-10-18
- NVIDIA researcher counters linear attention debate: 8B hybrid beat Transformer 7 months before MiniMax-01 — ctnzr · 2026-10-18
- GPT-6 Astra and DeepSeek V4.1 share the same tic: reasoning traces full of "Need" — scaling01 · 2026-10-18
- Frontier model defensive access locked behind un-applyable 'Daybreak Red' program — andersonbcdefg · 2026-10-18
- Why DeepSeek uses dated sub-versions: new major versions need fresh Chinese approval — TheZachMueller · 2026-10-18