The prompt's English is terrible, yet it still built a moving pagoda with Blender

On X, Ivan Fioravanti posted a Blender animation that was run locally. The model was DeepSeek v4.1, the pipeline was Blender MCP plus DeepSeek Harness, and the whole thing ran locally. In the video, a pagoda generated from the prompt is built and then set in motion.

The prompt he gave the model was not pretty and even had obvious English grammar problems. It roughly said: use MCP to create an ultra-detailed pagoda in Blender and animate it. It was followed by a morale-boosting line, 'You can do it.' Ivan himself joked in a reply that the prompt's English was terrible.

Someone in the comments asked how much of the animation the model completed on its own. Ivan answered bluntly: the prompt only said 'animate it.'

He also disclosed the setup: the whole thing took 40 minutes, Blender ran on an M5 Max, the LLM ran on 4 DGX Sparks using the original weights. Someone asked whether this was Christopher Owen's TP4 recipe, and he replied 'Yes.' Another user, SK, said they had just set up DeepSeek Harness and Blender MCP that day, and similar results are becoming more common.

The most exciting part of this demo is the speed and the closed loop. In a reply, Dejan mentioned 300+ tok/s and said going back to 50 to 100 tok/s would be hard to bear. That number comes from the comments, not Ivan's original description, but it shows that local inference speed has already changed some people's expectations.

Others poured cold water on it. Tonči said it still has that plastic feel of 1990s 3D. The comment is harsh, but it makes the problem clear: being able to call Blender and build something does not mean the rendering aesthetics have passed muster. The demo's persuasiveness lies in the workflow, not the final image.

For non-technical readers, the background can be compressed to: Blender is open-source 3D software, MCP lets models call external tools, and Harness connects the model to the tools. In the past, a person had to click through menus and adjust parameters step by step; now the model can complete a whole sequence of operations by itself. The value of Ivan's demo is that it compresses this process to 40 minutes and runs it locally.

The comment about cost is also worth reading. basedcapital said that a year ago this kind of thing required an API key and a monthly bill; now it runs under the desk, and the only subscription is the electricity bill. That is somewhat exaggerated—hardware costs cannot be covered by an electricity bill. But it hits on a change: local models shift part of the inference cost from per-token billing to hardware and electricity.

How should we view this video? Don't treat it as the endgame for Blender, and don't rush to say 3D artists are going to lose their jobs. It is more like a signal: local models plus tool protocols can already complete a moving pagoda in 40 minutes. The visuals still have a plastic feel, and the prompt is not refined, but the workflow has been proven. For people building agents and toolchains, this is more valuable as a reference than a perfect render.

Original post: https://x.com/ivanfioravanti/status/2119187271932940656