Cerebras Tests Kimi K2.6 Model with 29x Speed Boost in Long-Text Tasks

ME News reports that on May 20 (UTC+8), according to monitoring by Beating, wafer-scale chip company Cerebras has launched the trillion-parameter large model Kimi K2.6 in enterprise testing, completely eliminating interconnect latency from traditional board-level communication by integrating chips directly across an entire 12-inch silicon wafer. Third-party evaluation firm Artificial Analysis measured its generation speed at 981 tokens/s—6.7 times faster than mainstream GPU cloud services. In long-text tasks with 10,000 input tokens and 500 output tokens, total response time dropped from 163.7 seconds via Kimi’s official API to just 5.6 seconds—a 29-fold improvement. Since model weights are distributed across multiple wafers while activation values are streamed, inter-layer communication runs entirely on the wafer’s on-chip network fabric, achieving a physical communication bandwidth over 200 times that of NVIDIA’s NVLink in the NVL72 architecture. Combined with distributed computing optimizations, Kimi K2.6 stores weights natively in 4-bit format with minimal loss, uses 16-bit floating-point numbers during computation to preserve precision, and employs customized operator kernels with speculative decoding to achieve real-time performance. (Source: BlockBeats)