Fastino Labs argues for American open weights
Fastino Labs is arguing that open-weight releases should become the American default, and it cites its own fine-tunes of an NVIDIA Nemotron checkpoint as evidence that smaller labs can build on shared models.
Fastino Labs has published a position piece, written by Mary Newhauser, arguing that the United States should make open-weight model releases the norm. The company says it was one of more than two hundred organizations that signed an industry letter in July calling open weights a foundation of American AI.
The post makes three claims. Open weights speed up progress because labs can build on each other's work. They widen access beyond the few organizations that can afford pre-training. They also produce a range of models, both large and small, so builders can pick one that fits the task.
As evidence, Fastino describes work with NVIDIA's Nemotron team. Using what it calls an advanced fine-tuning agent, it adapted an early-access Nemotron 3.5 Lightning checkpoint into two open-weight domain models, one for finance and one for healthcare. Fastino reports these results against the base model:
- FinQA execution accuracy: 15.86% → 59.23%
- MedCalc-Bench: 49.09% → 54.18%
The company says the work used no pre-training compute and no NVIDIA data, only the shared checkpoint. It also says it published the ten best fine-tuning recipes its agent found. The post does not include independent verification of these figures.
Fastino also points to Stanford's 2023 Alpaca project, which fine-tuned Meta's LLaMA 7B on 52,000 instruction examples for under $600 of compute. It cites Epoch AI's Capabilities Index as showing that the quality gap between open and closed models is closing quickly. Its policy ask is that permissive licenses, such as MIT, BSD and Apache, become standard for model weights, so that a model release means an open-weight release unless there is a strong reason otherwise.