The first thing that stands out is not the 3D scene, but the token counter. In an experiment posted on Reddit, GLM 5.3 and GLM 5.3 Flash built a luxury duplex penthouse in Blender using BlenderMCP, the bridge that turns calls from an LLM into operations on the scene. The author ran both models locally on rented machines, with Q4 quantization: Flash takes between 190 and 200 GB plus context headroom, while the full model requires between 450 and 470 GB. Managing them means four RTX PRO 6000 WS accelerators for Flash and six for the base variant.

The most interesting part is not just the raw power involved: two models with very different VRAM footprints deliver similar object counts and timing, but with sharply different token consumption. Flash placed 811 objects in 38 minutes and 52 seconds, across 43 turns, with 9 tool errors and about 36,000 output tokens. The full model generated 847 objects in 40 minutes and 43 seconds, across 42 turns, with 8 errors and 112,000 tokens. It also spent about 22 minutes reasoning before placing the first object, while Flash got to work almost immediately.

To get the scene, the prompt had to include real dimensions: a 20 by 13 meter footprint, a 2.9 meter main ceiling, a double-height volume of 9 by 8 meters rising to 6.2 meters, a mezzanine at 3.1 meters, a staircase with 17 treads and a 0.182 meter rise, a 20 by 4.5 meter terrace, curtain-wall mullions every 1.5 meters, and PBR parameters for glass (refractive index between 1.45 and 1.52), concrete, marble, metal and fabric. The author’s early vague attempts produced what he called ‘3D goo’ instead of a room. Camera work was handled separately by Claude Opus 5 so it wouldn’t inflate the token statistics.

The structural takeaway is not who ‘won’, but that cost per token becomes the operational discriminator. The full model consumed more than three times the output tokens, added only 36 more objects and got the double-height volume wrong: it built 9 by 4.5 meters while reporting it as 9 by 8, whereas Flash got it right. In a self-hosted context, that gap is not a curiosity: it translates into more VRAM occupied, longer compute time and, if you rent hardware, higher per-session costs.

This changes incentives for teams developing on-premise 3D agents. A ‘capable’ model is not enough; you need a model that knows when to stop and use the tools at the right moment. The base model’s reasoning, with 82,000 tokens spent before acting, did not add accuracy where it mattered. At the same time, Q4 quantization does not make these models ‘light’ in any absolute sense: four-to-six GPU workstations are still needed, shifting the problem from file size to infrastructure management.

There is also a signal about the interface layer. The need to specify rises, spacing and refractive indices suggests the prompt is no longer a creative description but a form of technical specification. For teams integrating LLMs into 3D pipelines, the key skill becomes translating design constraints into parameters the machine can read.

This is not a benchmark, as the author rightly notes. But it is one of those informal tests that says more than many papers: the future of local agentic AI is not only about capability numbers, but about how much VRAM is needed, how many tokens are consumed and how much of that consumption turns into correct actions.