Hugging Face to White House: Open 7B models now beat Claude 3.7 on code

Hugging Face just told the White House something that should make proprietary AI vendors nervous: a 7-billion-parameter open model with the right post-training recipe can outperform Anthropic’s Claude 3.7 on complex coding tasks. Their new OlympicCoder project is the latest in a string of data points—alongside AI2’s fully open OLMo 2 models matching o1-mini performance—showing that the performance gap between open and closed systems has essentially collapsed. And it’s collapsing faster and with fewer resources than most people predicted.

The company’s formal response to the White House AI Action Plan RFI isn’t just cheerleading for open-source. It’s a direct argument that the administration’s strategy needs to treat openness as a prerequisite for performance, not a nice-to-have. Their logic: every major AI breakthrough—attention mechanisms, transformers, cheaper post-training algorithms—was built on open research. The commercial APIs everyone’s paying for are standing on a mountain of publicly funded, openly published science. Hugging Face wants policymakers to recognize that continuing to invest in that foundation produces multiplier effects that closed development simply doesn’t.

Where things get interesting is the efficiency pitch. The response argues that smaller, purpose-built models running on edge devices aren’t just cheaper—they’re actually more reliable in high-stakes domains like healthcare where generalist models have a track record of unpredictable failures. It’s a subtle but sharp jab at the one-model-to-rule-them-all approach that companies like OpenAI and Google are selling. If you’re a hospital, you probably don’t need GPT-5. You need a small model you can audit, fine-tune on your own data, and run without an internet connection.

On security, Hugging Face is essentially saying what the cybersecurity community has been saying about open-source software for decades: transparency is a feature, not a bug. Fully open training data and procedures enable actual safety certification. Open-weight models that run in air-gapped environments solve information risk problems that cloud-only APIs can’t touch. The subtext is clear: if the White House is serious about securing AI in critical infrastructure, the path runs through openness—not away from it. Whether that argument lands in a political climate increasingly focused on controlling AI access is the real question.