ML Inference

Making LLMs run fast on consumer hardware. Native Swift implementations using MLX for Apple Silicon — speculative decoding, OCR models, reasoning systems.

Projects

• speculative-decoding — MLX-Swift port
• nano-reasoning — FastRL on Apple Silicon
• paddleocr-vl.swift — 0.9B doc parsing VLM
• masked — Masked diffusion LM

Hardware Targets

• Apple Silicon (MLX)
• GPU — WebGL2/WebGPU
• CPU — Rust simd optimization