Developer Builds Fully Local Parkour Simulation Using Quantized GLM Model on vLLM
A developer created a fully local parkour simulation powered by GLM 5.3 Flash quantized with NVFP4 precision running on two DGX Spark GPUs via vLLM. The setup achieves prefill speeds of approximately 1,500 tokens per second and decode speeds of 40 tokens per second over large context windows. This project demonstrates the feasibility of running complex, interactive 3D spatial and code-generation LLM workflows locally without relying on cloud services. It highlights how combining 4-bit quantization with multi-GPU tensor parallelism makes high-performance local AI execution practical. The deployment uses vLLM with Tensor Parallelism (TP=2) and integrates Claude Code as an agentic scaffold with a 260k token context size. The author reported strong vision capabilities and solid 3D spatial understanding alongside interactive response times, publishing the deployment recipe on GitHub.
## BACKGROUND
LLM inference is split into two phases: prefill, which parallel-processes input prompts to build key-value (KV) caches, and decode, which sequentially generates response tokens one by one. Techniques like NVFP4 quantization compress model weights into 4-bit floating-point format to save memory, while tensor parallelism divides individual model layers across multiple GPUs to speed up compute and memory bandwidth.