In Silicon Valley's fiercely competitive decision model landscape, the open-source Karpo Decision 31B has secured the top spot in the Decision Index evaluation of 300 models, beating Google's closed-source Gemini Decision 1 across multiple metrics. The full lineup, ranging from 0.8B to 31B, has been open-sourced. It uses a mechanism that computes probabilities for all options in a single forward inference pass, replacing the traditional token-by-token generation approach. The 31B flagship version achieves a median single-card latency of 94 milliseconds, a calibration error of 0.023, and a cost of $0.025 per thousand decisions. In real-world scenario tests, its speed far outpaces general-purpose large models, and it demonstrates outstanding generalization capability and resistance to leaderboard gaming—making it well-suited to support the high-frequency, millisecond-level decision-making needs of AI agents.
