- Moonshot released Kimi K3, a 2.8 trillion parameter multimodal model with 1M token context window and native vision, now available on Modal.
- Modal runs K3 at 460 tokens per second on release day, with day-zero support from Moonshot and vLLM, including token-based pricing and dedicated endpoints.
- K3 is the strongest open model on public intelligence indexes, fourth overall, using a mixture-of-experts architecture with 896 experts, 16 active per token.
- Innovations like Kimi Delta Attention and Attention Residuals improve scaling efficiency by 2.5x over K2, while quantization-aware training and rebalanced expert parallelism enable broad hardware support.
- Modal provides a custom DFlash speculator tuned to K3, achieving 360% faster interactivity (100 to 460 tokens/sec) and 88% higher throughput (800k to 1.5M TPM per GPU) on agentic workloads.
- K3 is available via Modal's Shared API with token-based pricing or as a dedicated Auto Endpoint, with $30 free compute per month.