Radar ·
tiny-llm: open-source course for building a vLLM-style inference engine
Four-week open-source course that teaches building a small vLLM-style inference engine around Qwen3, covering attention, KV cache, 4-bit weights, kernels and batching, running on Apple Silicon with MLX.
According to the Radar source dated 4 November 2026, tiny-llm is a four-week open-source course in which the student builds a small vLLM-style inference engine around the Qwen3 model. The content covers attention, KV cache, 4-bit quantized weights, C++ and Metal kernels, and batching. The description states that the practice is guided and measured with KV cache, quantized weights and batching, and can run in a notebook without an NVIDIA GPU, on Apple Silicon with MLX.
The material matters to teams who want to understand what happens inside an inference server, but it does not describe any Rota Nacional service or capability. To verify the content, consult the original project page indicated in the Radar source and check the number of weeks, topics and hardware requirements before adopting the course. If you use an AI assistant to study the material, do not paste internal code, customer data or personal identifiers into any model; replace them with placeholders first.
Original source: tiny-llm: curso open-source para construir um motor de inferência estilo vLLM