One prompt, Blender MCP: Qwen4-Max builds a ramen stall at ~1/4 of DeepSeek's token cost
teortaxesTex · x · 2026-11-12
Using an identical prompt with only Blender MCP tools (no assets, code-only), @superalesha had two models build a cozy Tokyo ramen-stall diorama in Blender:
- Qwen4-Max: one prompt, xhigh effort, produced a full scene; consumed 3.02M input tokens (2.85M cached) + 62K output.
- DeepSeek Harness: 11.96M input (11.59M cached) + 130K output — roughly 4x Qwen's usage — though the result also looked quite good.
Quoting it, teortaxesTex calls it a "ridiculous DSH win", noting that without DeepSeek's cache economics the gap might be unjustified. He also hopes Qwen4's current instability is just a preview-phase issue and that the GA release will be more stable.
More from coding & agent
- AI-Generated PRs Are Drowning Open Source: Merging Now Requires Knowing a Maintainer — dbreunig · 2026-11-12
- 133 GPT-6 agents spend 6 days tuning Stockfish NNUE weights, gaining +17.66 Elo in 256-game match — imjustnewatai · 2026-11-12
- sokrypton: AI models are still terrible at coding — terminals made them usable, wet labs could do the same for biology — yuvalpi · 2026-11-12
- sokrypton: AI got usable at coding via terminal iteration — wet labs could do the same for biology — sokrypton · 2026-11-12
- Building a mangrove forest simulator with a swarm of Carbon agents citing 45 papers — AI_Andrew · 2026-11-12
- Fumadocs 16.19 adds runtime variables and branching without leaking hidden branches — bclavie · 2026-11-12