I have written gemma3 inference in pure C (github.com)

A new CPU inference engine called gemma3.c has been developed from scratch for the Gemma 3 4B IT model, showcasing that modern large language models (LLMs) can operate independently of Python, PyTorch, or GPUs. Written entirely in C11 with zero external dependencies, this innovation allows LLMs to run in resource-constrained environments, enhancing accessibility for developers who prefer a lightweight solution. Key features include the full Gemma 3 architecture, support for hybrid attention mechanisms, a native tokenizer with a 262k vocabulary, and efficient memory management through memory-mapped weights. This development is significant for the AI/ML community as it highlights the viability of running sophisticated models solely on CPU with optimized performance. The engine supports interactive chat mode and offers a comprehensive command-line interface along with a library API. With a memory footprint of approximately 3 GB and the ability to generate tokens in real-time, gemma3.c opens new avenues for deploying AI without relying on traditional computing environments. This project, licensed under MIT, emphasizes a shift towards more flexible and efficient model deployment strategies in AI.