Cerebras published a 59-wafer mapping for Kimi K3 on October 22, explaining how a model far larger than one chip's memory could run across its processors. The worked example gives developers a way to examine the memory budget behind fast inference. It isn't an independently tested deployment result.
The calculation assumes four-bit weights, an eight-bit key-value cache and an average request size of 256K tokens, according to Cerebras's post. Those settings define the example's scope.
What the 59-wafer count describes
Cerebras assigns one dense-attention layer or two linear-attention layers to each wafer, arriving at 24 plus 35 wafers. Weights stay in place while intermediate results pass between stages, the company explains.
The hardware unit deserves attention. Cerebras's WSE-3 Turbo product page describes a wafer-sized processor with 900,000 compute cores. Its CS-4 system contains three such processors. A wafer count and a system count therefore describe different quantities.
There is also rounding in the model size. Moonshot's Kimi K3 model card lists 2.8 trillion total parameters, with 104 billion activated, and 93 layers. Its attention composition is 69 KDA layers plus 24 Gated MLA layers. That supports the layer split used in Cerebras's illustration while putting the three-trillion figure in context.
Moonshot lists a 1,048,576-token context window and MXFP4 weights with MXFP8 activations in the same specification. The model's maximum context and Cerebras's assumed average request length answer different questions. A capacity estimate for 256K requests doesn't establish capacity at the full window.
The capacity question buyers should ask
Cerebras says cache space on the dense-attention wafers limits concurrent requests. It also describes a tradeoff between smaller micro-batches for speed and larger ones for throughput in its serving explanation.
For a team evaluating this design, the useful follow-up is a workload-matched result: how many simultaneous requests can it sustain at the context length the application actually uses? MrKeyoor has not benchmarked this configuration. The next evidence to watch is a measured run with those conditions disclosed, so buyers can connect the memory calculation to the service they would receive.