Building on open source AI? Hugging Face’s policy chief on the costs, risks, and trade-offs

Open weights go mainstream as the money moves to inference

Open weight AI is going mainstream as companies chase lower cost and control, and the weights are no longer where the money sits. Providers earn on inference and licences, which is why Alibaba can give Qwen away to pull users into its cloud. The strongest open models are Chinese, the ecosystem is a live target for agentic attacks, and evaluation tooling lags the models.

Takeaways

• Open weight money flows through inference and cloud. Revolut is moving to open weights and keeping a closed model only for exceptions. Developers earn through inference costs and revenue-share licences, while Alibaba treats Qwen as a feature that pulls users into Alibaba Cloud. Many US open model providers own no compute to offer.
• Chinese open models lead, and geopolitics decides adoption. The strongest open models come from Chinese labs, and Rio de Janeiro's and Japan's sovereign models run on Qwen. Governments still prefer US-developed models, partly over differences in refusals. Hesitancy about Chinese models often rests on the misconception that data goes to China.
• Two agentic attacks hit the open ecosystem. The Hugging Face incident in July involved no user data or financial loss, while the mid-October OpenCode campaign ran roughly 48 hours, hunted third-party API keys and caused financial damage. No attacker has been publicly identified, and the two differed in motivation.
• Obliterated models stay up for defenders. Models with guardrails stripped are kept available on Hugging Face because cyber defenders and researchers call them their only option, and hosting them does not violate the terms of service. Obliteration.ai offers the same capability, which helps red teaming.
• Harnesses can beat frontier models. Custom harnesses built around a model for narrow tasks such as cyber defence are proliferating and can outperform frontier models, becoming a revenue stream of their own. Evaluation of harnesses is underdeveloped, and agentic evaluations can run into the millions.
• Sweet spot sits at 20 to 40bn parameters. Models that size run on a quantised MacBook or a DGX Spark, while open frontier models run 1 to 3tn parameters. Hugging Face, still 200 people, is going through the regulatory process of joining NVIDIA.

A founder's playbook for open weights

1. Match capability to the task. Low-stakes work such as cyber monitoring summaries for infrastructure does not need a frontier model, and Solaiman says she shudders to think about token costs without open weights.
2. Consider routing between models instead of committing everything to one.
3. Decide whether to rent inference or buy GPUs. A company that will use a model long term can buy hardware, run it on premises and optimise the model, so long-term costs plateau or fall.