X · @teortaxesTex
· X / Twitter
it just works
it just worksAbdur Rahim: DeepSeek's DSpark speculative-decoding drafters, benchmarked on a Mac 👇Native on Apple Silicon (MLX), lossless; identical output to the base model, ~1.4–1.6× faster and up to ~2× on code/mathRecent releases turned it into a local OpenAI-compatible server (with continuous