Msplat – 3D Gaussian Splatting training in ~90s on M4 Max, native Metal
3DGS training in 90 seconds on M4 via fused Metal kernels, no PyTorch overhead—unprecedented Apple Silicon story.

Custom Metal shaders deliver 1.87× speedup over FP16 baseline with Int8 matrix multiplication.
ML engineers optimizing inference on Apple Silicon
MLX · llama.cpp · Stable Diffusion optimizations
3DGS training in 90 seconds on M4 via fused Metal kernels, no PyTorch overhead—unprecedented Apple Silicon story.
Autonomous agent wrote custom Metal kernels boosting decode speed 42% over upstream llama.cpp.
Running LLM inference in pure GDScript and Vulkan shaders inside a game engine.
Fused int4 attention kernel on Metal keeps LLM speed constant as context grows.
E8 lattice geometry replaces attention—clever math, but TinyStories 0.37 loss needs context.
Meshless Voronoi discretization running realtime on Vulkan is genuinely clever.