Cerebras trillion-parameter models: how it serves them

What’s new, in 20 seconds

In an October 22, 2026 post, Cerebras explained how it fits a trillion-parameter model across many wafers, using Moonshot's Kimi K3 as the example: 35 wafers for 69 linear-attention layers and 24 for its other 24 layers, 59 WSE-3T wafers in all, each with 44 GB of on-chip SRAM.
Layers run as a pipeline: only activations pass between wafers, over RDMA on Ethernet at 'roughly 2 microseconds per hop.' The plan assumes 4-bit weights and an 8-bit KV cache, and Cerebras says the approach targets '1000-2000+ tokens per second.'
All figures are Cerebras's own. The post gives no price or measured speed for Kimi K3; Cerebras's docs list K3, not K3.1, only for enterprise Dedicated Inference. Artificial Analysis measured 981 tokens per second for Kimi K2.6 on a private Cerebras endpoint on May 6, 2026.

Contents

How does Cerebras serve trillion-parameter models on AI chips that hold only 44 GB each? In an engineering post published October 22, 2026, Cerebras says it splits a model's layers across many wafers and runs them as a pipeline. Its worked example is Moonshot AI's Kimi K3, which it says needs 59 of its WSE-3T wafers, and it says the approach targets "1000-2000+ tokens per second." Kimi K3 is the base of Kimi K3.1 Pro, released on October 23, 2026 and covered in our Kimi K3.1 Pro story.

This story is based on Cerebras's post, "How we deploy Trillion-parameter frontier models on Cerebras" by Sarah Chieng, its August 18, 2026 CS-4 announcement and Kimi K2.6 posts, its pricing page and inference docs, and Moonshot's Kimi K3 model card on Hugging Face, checked October 25 and re-checked October 26, 2026. Every performance figure here is vendor-reported.

The short version

The problem: "A single Cerebras WSE-3T wafer has 44 GB of on-chip memory," while "trillion-parameter frontier models need terabytes for their weights alone."
The answer: pipeline parallelism. Each wafer holds some layers; only activations travel between wafers.
Kimi K3 example: 35 + 24 = 59 wafers, with 4-bit weights and an 8-bit KV cache.
Speed: a target of 1,000 to 2,000+ tokens per second, not a published measurement for K3.
Not in the post: price, availability of Kimi K3 or K3.1 on Cerebras's cloud, or a GPU comparison with numbers.

How does Cerebras fit Kimi K3 on 59 wafers?

Cerebras says each Kimi K3 layer is about 16 GB at 4-bit precision. Two linear-attention layers fit on one 44 GB wafer, so its 69 linear-attention layers take 35 wafers. Each of the other 24 layers gets a wafer to itself, leaving the remaining memory for that layer's KV cache, the stored attention state that grows with context length.

Layer type (Cerebras's terms)	Layers	Layers per wafer	Wafers
Linear-attention layers	69	2	35
"Dense" layers, with KV cache	24	1	24
Total	93		59

Table 1Most wafers hold two layers; the 24 attention-heavy layers each get one, to leave room for their KV cache. Source: Cerebras, "How we deploy Trillion-parameter frontier models on Cerebras" (October 22, 2026).

Moonshot's model card describes the same 93 layers differently: 69 Kimi Delta Attention (KDA) layers and 24 Gated MLA layers, with one dense layer, in a mixture-of-experts model with 896 experts. Cerebras's "24 dense layers" appear to be the 24 MLA layers, which keep a full KV cache. The card also lists the size as 2.8 trillion total and 104 billion active parameters, where Cerebras says "~3 trillion," and says K3 ships with native 4-bit MXFP4 weights, which matches Cerebras's 4-bit assumption.

By 9to5AI's arithmetic, 59 wafers at 44 GB each hold about 2.6 TB of on-chip memory. Cerebras doesn't say how many systems that is. Its CS-4, announced August 18, 2026, is built from three WSE-3T wafers, so a 59-wafer deployment would span at least 20 CS-4 systems; that's an illustrative calculation, not a Cerebras figure.

How does pipelining keep it fast?

Cerebras says it "executes models in a pipeline." Each wafer works on one stage, and "only activations need to be passed between the wafers," which it does over RDMA on Ethernet at "roughly 2 microseconds per hop." Because each stage can work on a different request, and each stage can process "a micro-batch, a small group of requests," the number of concurrent requests scales with the number of pipeline stages times the micro-batch size. The post's worked example assumes a 256,000-token average request.

The post doesn't mention tensor or expert parallelism. It contrasts its approach with GPU batching and argues the fair metric is "total token throughput per unit of power (e.g. tokens per second per megawatt)," without giving that number for Cerebras or any GPU system.

How fast is it, and what does it cost?

The post states a goal, "to run multi-trillion parameter models at 1000-2000+ tokens per second," and doesn't report a measured speed for Kimi K3. The nearest published figure is for an earlier Moonshot model: Cerebras says Artificial Analysis measured 981 output tokens per second for Kimi K2.6 on a private Cerebras endpoint on May 6, 2026, offered as enterprise trials.

No price is listed. The post points readers to a free API at cloud.cerebras.ai and to sales. Checked October 26, 2026, neither Cerebras's pricing page nor the model list for its shared, pay-as-you-go service names a Kimi model. Its Dedicated Inference documentation, for enterprise customers on reserved capacity, lists Kimi K3 among the models it supports, with no price; it doesn't list K3.1, and the post doesn't say Kimi K3.1 Pro is available on Cerebras.

For the CS-4, Cerebras claims "up to 30 times faster than GPU solutions" on GPT-OSS-120B at more than 4,400 tokens per second per user, without naming the GPUs; its footnote says throughput "varies by model architecture, context length, precision, and serving configuration." The Register reported that the WSE-3T keeps the WSE-3's 44 GB of SRAM and doubles its rated compute and memory bandwidth by running at higher power.

Why does it matter?

Memory is the constraint every inference chip designer is trading around. Cerebras keeps weights in fast on-wafer SRAM and spreads a model across dozens of wafers; Intel's inference GPU goes the other way, trading HBM for up to 480 GB of cheaper LPDDR5X, as our Intel GPU D270 story explains. Software matters too: our TorchTPU vs TorchNeuron comparison covers how PyTorch runs on Google's and AWS's own chips.

What's confirmed, reported and uncertain

Vendor-reported (Cerebras): the 44 GB per wafer, the 59-wafer plan, ~16 GB per layer, the 2-microsecond hop, the 1,000–2,000+ tokens-per-second target and the CS-4 claims.
Confirmed (Moonshot): Kimi K3's 93 layers (69 KDA, 24 Gated MLA), 2.8 trillion parameters and native 4-bit weights.
Listed (Cerebras docs): Kimi K3 among the models supported on enterprise Dedicated Inference, checked October 26, 2026.
Uncertain: real Kimi K3 or K3.1 speed on Cerebras, pricing, whether any customer runs K3 there yet, K3.1 availability, and how many systems a deployment uses.