vladimirrott / mojo-dllm

Why you might care

Runs diffusion LLMs without Python at runtime, with the GGUF reader, tokenizer, quantized matmul, GPU kernels and denoising loop all in Mojo. Its README benchmarks show it ahead of llama.cpp's diffusion CLI on the CPU run it tested (1.33 vs 0.69 tokens/s), and GPU is optional.