Research|LLM: Kimi K3 - Scaling Still Works; An Expensive Model Competing at Front Tier

Kimi K3's overall performance now sits in the global top tier. It is a large 2.8-trillion-parameter MoE model with a 1-million-token context window, and it stands out in tests spanning coding, agents, and long-horizon knowledge work. The first round of global user feedback has also broadly confirmed its capabilities, especially in complex coding and long-running task execution. However, the model has only just been released, and its real-world stability, success rates, and per-task cost still need further validation.

K3 shows once again that scaling laws have not broken down - larger models still generally deliver better absolute performance. Architectural optimization, MoE, distillation, and post-training can raise the efficiency with which parameters and compute are used, but for further gains in complex reasoning, long-horizon planning, and agent capability, increasing model size remains the most direct path. After K3 again scaled up substantially from the K2 series, its capabilities took a clear step up; this suggests architectural innovation is mostly improving scaling efficiency rather than replacing scaling itself.

Sentiment in the global community has been broadly positive. Many users credit K3's benchmark, coding, and long-horizon agent performance, while also worrying that it currently runs only at the highest level of reasoning effort, with no low-cost, low-latency, lightweight mode. In other words, K3 may be well-suited to high-value, complex tasks but not necessarily economical for large volumes of simple requests: the model may think longer and generate more reasoning tokens, so the inference compute required to complete the same task can exceed that of models that can dynamically adjust their thinking depth.