Cerebras Says Kimi K3 Fits on 59 Wafers, and Gives No Price for Any of Them

Cerebras published a walkthrough on Oct. 22 saying Moonshot AI's Kimi K3 model would occupy 59 of its WSE-3T wafer-scale chips, and that such deployments can serve multi-trillion-parameter models at "1000-2000+ tokens per second". The company gives no price, no customer and no measured run of K3.

The post, "How we deploy Trillion-parameter frontier models on Cerebras", is credited to Sarah Chieng. It calls K3 an open model of about 3 trillion parameters. The post labels the K3 sizing an illustrative example built on assumed settings: 4-bit weights, an 8-bit key-value cache and a 256,000-token average request.

How the wafer count is built

The post gives K3 93 layers, 69 with linear attention and 24 dense. At 4-bit precision the weights come to about 1.5 terabytes, or about 16 GB per layer. Each wafer carries 44 GB of on-chip memory.

Two linear-attention layers fit on one wafer, at 32 GB. That puts the 69 linear layers on 35 wafers. Each dense layer takes a wafer to itself, which leaves roughly 28 GB per wafer for the key-value cache. Cerebras totals the two groups at 59 wafers.

Layer type	Layers	Layers per wafer	Wafers
Linear attention	69	2	35
Dense	24	1	24
Total	93		59

Activations pass between wafers over RDMA over Ethernet at roughly 2 microseconds per hop, the post says. Weights stay put, and each wafer works on a different request at once. The 59 wafers hold about 2.6 terabytes of on-chip memory in total, by the arithmetic of 59 times 44 GB. That figure is this publication's multiplication, not Cerebras's.

What the model is, and what its licence allows

Cerebras's "about 3 trillion" is rounded up. VentureBeat, reporting on the weights release, gives 2.8 trillion total parameters in a mixture-of-experts design, 104 billion active per token, 896 experts and a 1M-token context. Moonshot released the full weights on July 27, 2026, according to VentureBeat's Carl Franzen.

The licence is not MIT. A company running a "Model as a Service" business needs a separate agreement with Moonshot once aggregate revenue, affiliates included, passes $20 million over any 12 consecutive months. Products with more than 100 million monthly active users, or more than $20 million in monthly revenue, must display "Kimi K3" on the interface. VentureBeat quotes Nathan Lambert describing the terms as "inspired by MIT but distinctly non-commercial".

That matters for the Cerebras pitch. A cloud selling K3 tokens is a Model as a Service business.

The benchmark numbers belong to Moonshot

The Cerebras post quotes Moonshot's own results: 88.3 percent on Terminal-Bench 2.1 against 88.8 percent for GPT-5.6 Sol, and 91.2 percent on BrowseComp against 90.4 percent. Cerebras says it did not verify them. The site's earlier report on GPT-6 models reaching Google Cloud covers the closed models K3 is measured against.

1. Linear-attention layers35 wafers
2. Dense layers24 wafers

Source: Cerebras blog post, Oct. 22, 2026 (vendor-supplied), accessed 2026-10-25

What the post leaves out

The speed claim is a range for "multi-trillion-parameter models", not a number for K3. The post says gains over GPUs vary by workload, configuration, date and model, and that comparisons come from third-party benchmarking or internal testing. It names neither.

It proposes tokens per second per megawatt as the efficiency measure but publishes no such figure. Concurrency depends on the key-value cache left on the dense wafers, and the post shows those equations only as images.

VentureBeat reports K3 has run on clusters of consumer RTX 5090 GPUs, without saying how many. The cost of 59 wafers is the missing comparison, and Cerebras has not published it. The site's coverage of a Linux module lifting the BF16 cap on GeForce RTX 3000 cards shows how much those consumer cards can differ from their spec sheets.

Cerebras's page points readers to its inference cloud, a free API and a sales contact.