A 35B model targets the real bottleneck in AI agents: cost
Long-horizon agents do not just need smarter answers; they need to track state, recover from mistakes, call tools, and finish work without burning through inference budget. Occamy-1.0 is an open 35B model trained on execution-grounded trajectories for coding, tool use, and file manipulation, and its evaluation places it at the low-cost knee of the measured cost-performance frontier. The weights and part of the training data are released, giving agent builders a practical smaller-model baseline to test.