In Cerebras’s Kimi K3 example, model deployment means placing the model’s layers on WSE-3T wafers and passing activations between them. Cerebras’s October 22, 2026 explanation assigns the model’s 69 linear-attention and 24 dense layers across 59 wafers, using specific assumptions for weight precision, cache and request size.
That allocation turns a model described as roughly 3 trillion parameters into a set of memory-sized stages. The wafer count and request estimate below belong to Cerebras’s illustrative configuration.
Why the example assigns 59 wafers
Cerebras states that each WSE-3T wafer has 44 GB of on-chip SRAM, or static random-access memory. Its Kimi K3 example assumes 4-bit weights, an 8-bit key-value (KV) cache and an average request size of 256k tokens. Cerebras estimates the model’s raw weight payload at 1.5 terabytes, or about 16 GB per layer across its 93 layers.
| Layer type | Kimi K3 layers | Layers per wafer | Wafer count | Memory assignment |
| Linear attention | 69 | 2 | 35 | About 16 GB per layer; two layers assigned to each wafer |
| Dense | 24 | 1 | 24 | About 16 GB for one layer, with about 28 GB remaining for KV cache |
The arithmetic is 35 wafers for the linear-attention layers plus 24 for the dense layers, totaling 59. Cerebras’s placement rule fits two roughly 16 GB linear-attention layers on a wafer. It assigns one dense layer to each wafer, leaving the remaining on-chip memory for KV cache.
How requests move through the wafer pipeline
Cerebras describes the model weights as staying on their assigned wafer stages while activations move between stages. The company says that transfer uses RDMA, or remote direct memory access, over Ethernet, at roughly 2 microseconds per hop.
With requests at different stages at the same time, the pipeline can work on multiple requests in parallel. Cerebras also describes using micro-batches, or groups of requests processed together: smaller batches favor speed for an individual request, while larger batches favor total throughput.
What the 190-request estimate represents
For its dense-layer wafer calculation, Cerebras estimates about 28 GB of remaining cache memory and about 147 MB of cache for each assumed 256k-token request. Its graphic uses those rounded figures to estimate roughly 190 concurrent requests. That estimate is tied to the stated cache and request assumptions.
How AI SDK access differs from wafer deployment
The physical mapping assigns model layers to wafers; an application-level SDK provides a separate route for calling a Cerebras model. The AI SDK documentation for Cerebras describes both Gateway routing and direct provider access.
In Gateway, order: ['cerebras'] expresses a preference. A request may fall back to another available provider if Cerebras is unavailable or the request fails. For provider-specific APIs or custom connection settings, the documentation describes direct access through @ai-sdk/cerebras, configured with a CEREBRAS_API_KEY.
