Moonshot AI Unveils Kimi K3.1 Pro AI Model Capable of Processing 1 Million Tokens

    Chinese artificial intelligence startup Moonshot AI unveiled its Kimi K3.1 Pro model on October 24. The new model has approximately 3 trillion parameters and can perform programming tasks. The company deliberately weakened some of the model’s capabilities to make it more difficult to use in cyberattacks. The model’s core files will be made available to developers on November 4.    Moonshot AI removed material related to cyberattacks from the model’s training data. The goal is to preserve its programming capabilities while weakening its ability to perform tasks that could be used for attacks. In the company’s internal testing, Kimi K3.1 Pro scored 81% on a cybersecurity test. A similar model without the same restrictions scored 90.5%.    The safety measures therefore reduced the score by 9.5 percentage points. However, the test has not been verified by independent parties. Nevertheless, the new model outperformed the previous version. Its programming capabilities could help defenders find vulnerabilities in software. Attackers could also use the same capabilities to write malicious code and automate multistep tasks.    Kimi K3.1 Pro can process 1 million tokens at a time. A token is a small piece of text processed by artificial intelligence. This limit means the model can analyze a large volume of information at once. Moonshot AI also demonstrated a Highspeed version that is said to operate 2-3 times faster than the standard model. The company has not released independent speed test results.    China has also proposed imposing export restrictions on certain technologies used to develop malware with artificial intelligence and find software vulnerabilities. The proposals are still under public discussion. Once Kimi K3.1 Pro’s files are released, developers will be able to run and modify the model on their own systems. When that happens, Moonshot AI’s control over the safety restrictions will be significantly reduced.