Kimi has officially released the K3.1 Pro model. As the first Pro version in the K3 series, its key breakthroughs lie in task-execution capabilities. In the Terminal-Bench 4.0 test conducted by the independent evaluation firm Artificial Analysis, the K3.1 Pro’s completion rate surged from 12.6% for the K3 to 67.2%, placing it at the top of the current ranking; this test requires models to fully execute commands and pass all validation checks in a real-world computer-terminal environment. Meanwhile, on the AutomationBench-AA benchmark—covering cross‑email, spreadsheet, and CRM office tasks—the target completion rate rose from 58.3% to 73.3%, while GDPval-AA—focused on document, spreadsheet, and presentation generation—ranked third, demonstrating enhanced practicality in real‑world productivity scenarios.
Performance improvements come with adjustments to the technical roadmap: While the official release does not disclose specific optimizations on the inference side, it emphasizes increased post-training investment and the use of joint reinforcement learning across hundreds of downstream domains to enhance long-context processing and task-execution environments. Notably, the gains exhibit structural differences: in the HLE full‑scale benchmark, Pro scores 70.0% when the tool is enabled (higher than Opus’s 5.5%), but drops to just 51.1% when the tool is disabled (lower than Opus’s 59.8%); in the AA long‑context reasoning test, K3 (Max) achieves 88.7%, while Pro scores 86.3%, suggesting that its advantages are more concentrated in tool‑invoked closed‑loop execution rather than purely text‑based deep reasoning.
The most signal‑rich demonstration was the official human‑walking simulation: K3.1 Pro took full control of the controller training. When the virtual human’s arm movements remained persistently abnormal, the model did not stop at tweaking the reward function; instead, it proactively inspected the simulator’s code, identified and fixed 36 errors in the calculation of shoulder muscle force arms, and then restarted training. This “diagnosis–repair–retraining” closed loop surpasses the traditional RLHF paradigm, showcasing an understanding of and ability to intervene in the intrinsic logic of physics‑based simulation systems—widely regarded as a nascent hallmark of embodied intelligence research.
The commercial strategy has shifted to support high-frequency interactions: the standard Pro‑version API pricing has been reduced to RMB 68 per million tokens (RMB 100 for K3), while the Highspeed version costs RMB 136 and only accelerates the model’s output stage; latency in file I/O, command execution, and other stages remains unchanged. User testing shows that complex tasks—such as 3D controller modeling or a pelican riding a bicycle animation—often trigger quota bottlenecks after multiple rounds of iteration. In particular, users on the RMB 199 plan exhaust their allowance after just three rounds of modifications, whereas the RMB 599 plan exhibits a markedly different consumption rate, highlighting that the current token‑metering rules do not yet provide clear visibility across the entire workflow of long‑running tasks (generation → validation → modification → export).
Based on overall technological trends and organizational moves, Kimi’s embodied intelligence strategy has transitioned from concept to practice: in addition to the aforementioned simulation capabilities, public reports confirm that Moonshot is actively recruiting talent in the fields of embodied intelligence and robotics, with some candidates coming from leading robotics companies. On the technical front, its approach echoes Google Gemini Robotics ER 2—where the latter employs a Flash model for environmental perception and task planning, while a lower-level motion model executes actions—and the simulation design and diagnostic capabilities demonstrated by K3.1 Pro are ideally suited to serve as the “agent brain” on the training side. If a low-latency, low-cost Flash variant is released in the future, it could complete the real-time perception–planning–execution closed loop.