Open-source models remain limited to simple agent sub-tasks

Harrison Chase stated that open-source and post-trained models are primarily deployed for simpler tasks such as routing, guardrails, extraction, and summarization. He noted that these models do not yet drive the core agentic loops of complex systems, where frontier models remain dominant. Companies like Harvey and Ramp are experimenting with post-training open-weight models, often based on Chinese architectures, for these specific sub-tasks. Chase acknowledged that current benchmarks show some lag behind Chinese models but believes comparable US open-source models will emerge within the next 12 months. He cited NVIDIA’s ecosystem support and recent launches by Reflection as key drivers for this shift. Regarding data privacy, Chase argued that fears of using Chinese open-source models in enterprises are largely unfounded if the models are self-hosted or accessed via providers like Fireworks. However, he acknowledged potential concerns regarding value alignment in specific domains. He predicted that post-training will become more common for high-volume, simple components of agent systems, while the core reasoning loop will continue to rely on the latest frontier models until open models catch up in capability.