Show HN: LingBot-World 2.0 (1.3B) running at 16 FPS on one RTX 5090

Category: ai-ml

Tags: world-model, inference-optimization, pytorch, fp8, gpu-kernels

Score: 6.5/10 (Innovation: 5, Technical: 8, Documentation: 8, Utility: 5)

A 1.3B-parameter world model inference fork that runs at 16 FPS on a single RTX 5090, a 2.7x speedup over the original paper's code with lossless output. The author systematically optimizes the full stack—FP8 matmuls, SageAttention, fused kernels, and a compiled fp16 decoder—documenting each step with measured results.

Target audience: ML engineers and GPU performance engineers working on real-time generative models

Repository: https://github.com/kaarelkaarelson/lingbot-world-v2-realtime · Python · NOASSERTION · 15 stars

View on Hacker News