tiny-llm
A hands-on course for systems engineers that builds a small LLM inference and serving stack from scratch on Apple Silicon, following a Qwen3 model from array operations through a miniature vLLM.
tiny-llm is a course for systems engineers who want to understand LLM inference end to end, built on MLX arrays and the MLX extension runtime rather than high-level neural-network layers. Learners implement each operator themselves in Python, C++ or Metal, using MLX's own optimized version as the correctness and performance reference.
The course follows a four-week path: building a Qwen3 model directly from array operations in week one; adding a KV cache, quantized weights and faster prefill in week two; building a miniature vLLM with continuous batching and paged KV cache in week three; and building a bounded, checkpointed coding-agent loop in week four.