Hugging Face's two-person squad behind Zephyr

Hugging Face's H4 project gives the open-source community complete recipes and code to build ChatGPT-style conversational models using accessible 7B architectures.

Behind the work is an ultra-lean two-person research unit formed in January 2023. Dubbed H4—short for helpful, honest, harmless, and huggy—the team was assembled right after ChatGPT exploded, specifically to figure out how open-source libraries and models could replicate frontier chat performance.

How two engineers built Zephyr

The team's headline release is Zephyr-7B-α, a fine-tuned chat variant based on Mistral 7B. H4 also modified the Falcon-40B base model from Abu Dhabi's Technology Innovation Institute, tuning it to answer everyday natural language prompts much more effectively.

The engineers driving it, Lewis Tunstall and Ed Beeching, work remotely across Europe. While the core team is intentionally kept to two people for agility, they lean on a massive internal hardware setup: a dedicated cluster packing over 1,000 Nvidia A100 GPUs.

They also lean on internal testing units and collaborate externally with open-source stalwarts like LMSYS and LlamaIndex to benchmark and validate their releases.

What is open and where to find it

H4 is not just dumping weights and walking away. This month, the duo published an open-source handbook containing all the training source code and datasets used to create Zephyr.

Their main research agenda centers on alignment—testing whether community and industry fine-tuning ideas actually hold up in practice. The handbook will be continuously updated with recipes from future models as they ship.

While H4 does not directly monetize its code, the research feeds directly into Hugging Face's enterprise-focused Expert Acceleration Program, helping paying clients build custom internal AI setups.

Why it matters

H4 shows that the bottleneck in open-source chat models is no longer model architecture, but transparent alignment recipes. By publishing the full pipeline alongside Zephyr-7B-α, Tunstall and Beeching have given builders a proven playbook for turning raw open-weight models into capable assistants without reinventing the wheel.

While a 1,000-A100 cluster remains out of reach for independent developers, open-sourcing the exact data and training code democratizes the downstream fine-tuning process. It sets a welcome precedent: open source does not just mean open weights, it means open instructions.