Show HN: DeepSeek V3 in ~200 lines of pure Jax

A developer implemented the DeepSeek-V3 language model architecture in approximately 190 lines of pure JAX code, training it from scratch on a MacBook Air using public-domain detective fiction. The implementation demonstrates key features from the DeepSeek-V3 paper including Multi-head Latent Attention, Decoupled RoPE, and DeepSeekMoE routing, achieving a validation loss of 3.51 and token generation speeds of ~10.7k tokens/second.