October 12, 2026 – Moonshot AI’s Kimi K3, currently regarded as the strongest domestic large model in China, is approaching the three-month mark since its release. Rumors of an upgraded K3.1 have circulated since late last month, with reports now pointing to a mid-October launch.
According to Nomura Securities, Kimi K3.1 is expected to be released before mid-October, with a focus on improving the model’s ability to continuously execute long-duration agent tasks. The team plans to further optimize the model through post-training, using computer operation trajectories and real-world task data.
The current Kimi lineup is divided into two tracks. Flagship models such as K3 and K3.1 continue to push the upper bound of intelligence, while Kimi K2.8, released on September 11, emphasizes efficiency for coding and everyday agent workloads.
Beyond K3, Moonshot AI is also training a more powerful model, reportedly targeting 5 to 6 trillion parameters, widely believed to be the previously rumored Kimi K4.
For Kimi, the main constraint remains compute capacity, especially access to high-end training hardware. Domestic AI chips are generally adequate for daily inference, but large-scale training still cannot fully replace Nvidia GPUs in the short term.
Kimi K3 had already set a new performance benchmark for domestic models when it launched. With 2.8 trillion parameters, it was the largest open-weight model in China and among the largest globally, and it was seen as the first domestic model to truly approach frontier-level performance. Its significance grew after DeepSeek V4 Pro’s official release fell short of expectations, leaving K3 as a leading domestic performance reference.
However, K3 has faced several issues over the past three months. Availability has been limited at times, with some subscription plans temporarily restricted; conditions have eased slightly, but not enough to fully satisfy users. Pricing has also become a concern. Although Moonshot AI has argued that domestic AI models do not necessarily need to be cheap, K3’s cost advantage has weakened as OpenAI and Anthropic recently launched new products alongside significant price cuts.
Kimi also lacks a truly competitive Flash-tier model. K2.8 Preview is positioned for daily use, but it has not yet reached the level of DeepSeek V4.1 Flash, which combines strong performance, speed, and low cost. In that segment, Kimi still lacks a flagship product.
The upcoming Kimi K3.1 is expected to address some of these gaps, particularly in agent reliability and long-task execution. Whether it can also improve availability and cost competitiveness will be key, especially as OpenAI and Anthropic continue to raise the bar while domestic teams still face tight high-end compute supply.
