Kimi Open Source Model K3 Nears GPT-5.6 Sol and Fable 5, but Prices Surge

Kimi launched open-weight model K3, a multimodal model with 2.8 trillion parameters and 1M token context. In Kimi's benchmarks, K3 approaches Claude Fable 5 and GPT 5.6 Sol, and significantly outperforms Opus 4.8 and GLM 5.2 on many tests. However, K3's price has risen sharply compared to its predecessors, potentially shifting expectations for cost advantage of domestic models. Full weights are scheduled for release on July 27. For teams planning to deploy open-source models, K3's high performance is enticing, but its high cost makes it unsuitable for budget-sensitive scenarios. Developers needing top-tier capabilities may pivot to closed-source APIs or higher-cost local deployment; those seeking low cost may need to wait for other domestic open-source models.