Teaching Everyone to Fish for Tokens

Open-source language models are often compared to foundational software like Linux, but true open-source models require full training recipes, code, and data. In contrast, open-weight models only provide the model weights and inference code, functioning more like specific installed software versions. While open-weight models have a long shelf life, the full open-source recipe is a resource-intensive process that allows anyone to produce new weights and contribute community improvements.

Nvidia is investing heavily in nearly open-source models like Nemotron, releasing available data and training code to encourage countless builders and create massive demand for its inference chips. However, open-source AI faces an existential financial window because training models is extremely capital intensive. The ecosystem may rely on Nvidia’s financing to prove profitable, or open models may shift away from leading closed models to focus instead on efficiency, modifiability, and specialized enterprise-specific agents.

At the same time, training is becoming increasingly complex and abstracted, shifting away from full model training toward post-training and reasoning training. This trend reduces the number of builders creating base models and coincides with experiments in revenue-share licenses to keep near-frontier open weights viable. Meanwhile, companies like Meta use open-weight releases to commoditize competitors and hamper the revenue growth of companies relying on API token sales.