Large Language Models
Oct 25, 2026
Cerebras Explains Deployment of Trillion-Parameter AI Models on WSE-3T Chips
Oct 25, 2026
AI Summary
Cerebras has detailed its method for deploying large AI models, specifically a 3 trillion-parameter model named Kimi K3, across multiple WSE-3T wafers. This approach allows for efficient processing of requests at high speeds, utilizing the chip's on-chip memory effectively.
- Cerebras uses WSE-3T wafers, each with 44 GB of on-chip memory, to deploy large AI models.
- The Kimi K3 model has approximately 3 trillion parameters and is structured with 69 linear-attention layers and 24 dense layers.
- Kimi K3 achieved an 88.3% score on Terminal-Bench 2.1 and a 91.2% score on BrowseComp, demonstrating competitive performance against leading models.
- To run Kimi K3, a total of 59 wafers are required, with each layer needing about 16 GB of memory.
- Cerebras employs a pipelined execution model where each wafer processes different layers and requests simultaneously, enhancing throughput.
- The system can handle multiple requests at once, with the number of concurrent requests determined by the on-chip memory allocated for KV cache.
- The efficiency of the inference system is measured by total token throughput per unit of power, allowing for scalable deployment of larger models across multiple wafers.
trillion-parametercerebrasai modelsdeploymentmachine learning