Open mixture-of-experts LLMs, code generation, long context, linear attention (Gated DeltaNet), training on free compute (JAX/TPU)