Cerebras Runs Kimi K2.6 at 981 Tokens/Second — 6.7x Faster Than Any GPU Cloud on a 1T-Parameter Model
Cerebras published inference benchmarks for Kimi K2.6 — Moonshot AI’s 1-trillion-parameter open-weight model — running on its wafer-scale silicon. The headline number: 981 tokens per second, 6.7x faster than the next fastest GPU cloud option. Artificial Analysis validated the result independently.
Why GPU Clusters Struggle at This Scale
A 1-trillion-parameter model cannot fit on a single GPU. GPU-based inference systems shard the model across hundreds of accelerators and pass activations and weights between chips over high-speed interconnects — typically NVLink or InfiniBand. That cross-chip communication dominates latency at scale: the math completes quickly, but the data movement does not.
Cerebras uses wafer-scale chips, where the entire model resides on one physical processor. The routing between compute elements stays on-chip with dramatically higher bandwidth and lower latency than any external interconnect. For Kimi K2.6 at 1T parameters, that architectural difference translates directly into throughput.
What 981 Tokens/Second Actually Means for Agents
Raw token throughput matters differently for different use cases. For chat, a 6.7x speed gap is noticeable but rarely decisive. For agentic coding workflows, the arithmetic changes.
A typical coding agent loop — generate code, run tests, parse errors, reason about fixes, retry — requires dozens to hundreds of inference calls. Each call on a GPU cluster at ~145 tokens/second takes meaningfully longer than the same call at 981 tokens/second. Across a multi-step debugging or refactoring session, latency compounds: a task that takes 40 minutes on GPU infrastructure completes in under 6 minutes on Cerebras.
For enterprise teams running continuous CI-integrated agents, or for platforms billing by the session rather than by the token, the cost of waiting is real.
The Kimi K2.6 Context
Moonshot AI released Kimi K2.6 in late April — a 1-trillion-parameter mixture-of-experts model with 32B active parameters, open-weighted under a permissive license. It leads HLE-Full and SWE-bench Pro, beating every major proprietary model on both at release. The open-weight license made it a natural target for inference optimization from third-party providers.
Cerebras positions this as enterprise readiness for frontier open-weight models: 1T-parameter capability at a latency profile that closed-source API users associate with smaller, faster models.
Numbers
| Provider | Model | Throughput |
|---|---|---|
| Cerebras | Kimi K2.6 (1T params) | 981 tokens/sec |
| Next GPU Cloud | Kimi K2.6 | ~145 tokens/sec |
| Speed Gap | — | 6.7x |
Source: Cerebras blog, validated by Artificial Analysis.
The 6.7x claim covers the gap between Cerebras and the fastest available GPU cloud option for Kimi K2.6 at time of publication. Individual results vary by batch size and load.