Kimi K3, a new large language model, appeared on the LMArena benchmark platform under the label 'Kivine,' with users identifying the system on July 15 and sharing output samples. The discovery includes assessments indicating the model supports a context window of one million tokens and generates high-quality responses on long-horizon tasks, positioning it to compete with frontier systems.

Early tests revealed a notable trade-off in performance, with one user noting the model was significantly slower than competitors, taking 35 minutes to complete a single output generation task. The activity coincides with a brief promotion for the model that appeared on the company's documentation site earlier in the week, offering discounted top-up bonuses.