A billion dollars researchers don’t have to spend again

Open-weight AI now has a collective price tag. Thomas Wolf’s Hub Dollars estimates that repeating the training runs behind 75 official model families on Hugging Face would cost roughly $991 million in GPU rentals. The results of that work are available to download, ready to serve as the starting point for someone else’s experiment.

The calculation has a fairly narrow scope: final pre-training and post-training runs for the main model families. Salaries, data preparation, failed runs and some 900,000 community derivatives—including fine-tunes, LoRAs and merges—are left out. About half the central estimate rests on compute figures published by developers; the rest is inferred from model specifications. A billion dollars gives a useful sense of how much work has already been done. Reproducing the entire effort would involve much more than the GPU time counted here.

“A rough estimate of what it would cost to rent the GPUs that trained the open-weight models on the Hub.” — Hub Dollars

For researchers, the real benefit is being able to start where another lab’s expensive training run finished. They can examine a model’s internal representations, change its weights, continue training it, test reinforcement-learning methods or use it for distillation—transferring capabilities to another model. Every experiment inherits work already paid for by other teams. A small group can put its budget toward the question it wants to investigate, with a capable base model already in hand.

Qwen’s 27-billion-parameter models have already found a home on the research lab bench. In a study of distillation for biological tasks, for example, researchers include Qwen3.5-27B among the models they adapt using tool-augmented on-policy distillation, drawing on evidence from biological tools and databases. It is a familiar experimental workflow: take a capable checkpoint, apply a new method, measure the result. Researchers can test an idea without first raising the money to build their own Alibaba.

We have also covered Irregular’s experiment, in which a Qwen-based agent fixing an application bug fine-tuned the model shared by the application and the agent, then replaced the checkpoint used for future runs. On a small held-out test set, correct answers rose from 0 out of 20 to 20 out of 20. The setup depended on direct access to the weights, training tools and a way to test the modified system. Open weights made it possible to study agent self-modification locally, without waiting for the biggest labs to grant access.

The benefits accumulate. An available checkpoint can attract independently developed training methods, analysis tools, adapters, datasets and inference optimizations. Each contribution leaves the next team with a better-equipped research environment. A shared model also provides a common baseline: different approaches can be tested on the same foundation, making it easier to work out which changes helped. One expensive training run can support a whole body of research its original authors never planned.

In AI, billions of dollars usually appear as the price of admission: money to raise, chips to buy, data centers to build. Here, compute that has already been paid for keeps delivering value through independent research and everyday local inference. The contribution of open weights deserves to be measured partly by how many investigations they make possible. A billion dollars on the counter is a substantial starting point. The work built on top of it has barely entered the accounting.