Local LLM built a Blender scene needing 49GB of VRAM on an RTX Spark laptop

A test on the HP OmniBook Ultra 16 with Nvidia RTX Spark (20-core Grace CPU, 6,144-core Blackwell GPU, 128GB LPDDR5X) rendered a Blender scene using about 49GB of GPU memory — more than an RTX 5090 or a 48GB RTX PRO 5000. The procedural city of 13,600 objects was generated by a local Qwen3.8 Flash Next model (125B parameters, MoE) in 81 minutes, with each 1080p frame taking about 12 minutes. The laptop costs $6,229.99.

• The Blender scene used ~49GB of GPU memory, more than the 32GB RTX 5090 or 48GB RTX PRO 5000
• Qwen3.8 Flash Next has 125B parameters and runs up to 45 tokens/s with a 262K context
• The model generated a 13,600-object city and 700 lines of Python in 81 minutes
• Each 1080p Cycles frame took about 12 minutes; the configuration costs $6,229.99