Cerebras chips run Kimi K2.6 at 1,000 tokens/sec, 6.7x faster than GPUs

Cerebras Systems announced on May 20, 2026, that its wafer-scale chips now power Moonshot AI's Kimi K2.6, a trillion-parameter open-weight Mixture-of-Experts model, achieving record inference speeds of nearly 1,000 tokens per second for enterprise customers. This performance is independently verified by Artificial Analysis as 6.7 times faster than leading GPU-based cloud providers. The speed gain is critical for latency-sensitive applications like real-time code generation and autonomous agent workflows, where Kimi K2.6 specifically excels due to its architecture optimized for coding and agentic tasks.

Moonshot AI, based in Beijing, released Kimi K2.6 on April 20, 2026, as an open-weight model under a permissive license, enabling enterprises to deploy it on Cerebras hardware without vendor lock-in. The combination of Cerebras' custom silicon and the MoE model's sparse activation pattern allows this throughput leap, making it a compelling alternative for organizations needing high-speed AI inference at scale, rivaling closed-source offerings from major US providers.