GLM-5.3-Flash 通过 MCP 控制 Blender 完成复杂建筑建模，成本仅 $0.0526，比 GLM-5.3 的 $0.8807 低 16.7 倍，且耗时更短（38m 52s vs 40m 43s）。Flash 版严格遵循提示词，而 GLM-5.3 出现偏差，且 Flash 可在 128GB MacBook 上本地运行。这表明许多智能体任务可能过度配置了模型智能。
Love this, GLM-5.3-Flash produced a comparable Blender result for 16.7x less than GLM-5.3.
atomic[.]chat (@atomic_chat_hq, a desktop app that runs LLMs locally) gave both models a constrained architectural specification through MCP, with the same prompt.
The models controlled Blender through MCP over ~40 minutes, while making a sequence of tool calls to create and modify 800+ objects until the task is complete.
Note, that GLM 5.3 Flash can be run on a 128GB MacBook.
Outputs:
GLM 5.3 Flash: 811 objects, 38m 52s, $0.0526
GLM 5.3: 847 objects, 40m 43s, $0.8807
A lot of agentic tasks might be overprovisioned on intelligence.
GLM 5.3 Flash performs at GLM 5.3 level in Blender for 17x cheaper! We gave both models a live Blender over MCP and one prompt: a 2,800 sq ft duplex penthouse, double-height living room, mezzanine, floating stair, curtain wall, terrace, furnished, real PBR materials Outputs: GLM 5.3 Flash: 811 objects, 38m 52s, $0.0526 GLM 5.3: 847 objects, 40m 43s, $0.8807 GLM 5.3's prompt following was worse and it ended up making the penthouse smaller in depth, meanwhile GLM 5.3 Flash followed the prompt without any issues, along with that the flash model can be run on a 128GB MacBook 💻在 X 查看被引用的帖子
来源：Rohan Paul · x.com