Abhishek Hasrshvardhan Mishra

Abhishek Harshvardhan Mishra, widely known as ‘Tokenbender’ in the open-source AI community, is an AI researcher and engineer focused on LLMs, coding models, and agentic systems. Formerly a technical lead at Intel, he led high-impact projects in the server firmware division before transitioning deeply into language modelling and open-source AI research.

Over the years, Mishra has built and fine-tuned several influential open-source models and datasets. He created codeCherryPop from Llama 2 7B, regarded as one of the earliest useful small coding models, and later developed evolvedSeeker, an upgraded version of DeepSeek Coder 1.3B that became one of the best-performing sub-7B local coding models. 

He was also among the top three on the open Hugging Face leaderboard with his PIC (Partner-in-Crime) model series, which showcased advanced function-calling and character engagement capabilities.

Mishra additionally built one of the first models optimised for both RAG and generic chat in Indic languages, while his Hugging Face models and datasets have collectively crossed hundreds of thousands of downloads. Currently, he is researching novel methods to extract and deploy capabilities from language models without incurring the full latency and compute costs of large systems.