Georgi Gerganov

Gerganov’s personal portfolio links the model runtimes alongside earlier experiments in audio, terminal interfaces and data transmission through sound. In February 2026, Hugging Face announced that he and the GGML team were joining it to support ggml and llama.cpp. The announcement says the team would continue maintaining the projects with technical autonomy and an open community, while Hugging Face supplied long-term resources and worked on integration and packaging.[1][2][3]

Running language models across hardware

llama.cpp provides a C/C++ runtime for language and vision-language models. It targets CPUs and multiple accelerator backends and supports quantization, storing model numbers with fewer bits to reduce memory needs. Its hybrid execution can split work between CPU and GPU. These mechanisms make local inference practical on a wider range of devices; they do not mean every model fits every machine or retains identical quality under every quantization.[4]

Speech recognition without a remote service

whisper.cpp implements inference for OpenAI’s Whisper speech-recognition models in C/C++. Its repository documents CPU execution, GPU backends and examples running fully offline on an iPhone. Gerganov’s contribution is the implementation and tooling around the model, while the underlying Whisper model originated at OpenAI. The small runtime can be embedded into applications that need transcription without sending each recording to a hosted API.[1][5]

The tensor library underneath

ggml supplies portable operations on tensors, the multidimensional arrays used by neural networks. Its optimized kernels, quantized numeric formats and hardware backends provide a common foundation for the higher-level runtimes. Separating that foundation from model-specific code lets community work on low-level execution improve more than one application.[4][5][6]