On October 22, 2026, Cerebras published a technical article on its official blog, authored by Sarah Chieng and titled "How We Deploy Trillion-Parameter Frontier Models on Cerebras." The article uses Moonshot AI's open-source model Kimi K3 as a case study and details a method for distributing and running a roughly 3-trillion-parameter model across 59 WSE-3T wafers, reaching inference speeds of 1,000 to over 2,000 tokens per second.
A single WSE-3T wafer has only 44 GB of on-chip SRAM, while the weights of a 3-trillion-parameter model require several TB of space. Cerebras splits the model and distributes it across multiple wafers for cooperative inference in a pipeline.
Kimi K3 has about 3 trillion parameters (multiple media reports put the exact number at 2.8 trillion), with 93 layers in total—69 linear attention layers and 24 dense attention layers. Stored with 4-bit quantization, all weights total about 1.5 TB, averaging about 16 GB per layer.
Dense attention layers need to store a KV cache for active requests, and memory usage grows linearly with concurrency. Cerebras's approach places one dense layer per wafer, so 24 layers correspond to 24 wafers. Linear attention layers maintain only a fixed-size state and require no additional KV cache; each wafer can hold two layers, so 69 layers require 35 wafers. The total is 59 WSE-3T wafers.
The 59 wafers use a pipelined execution architecture: the compiler maps model layers to each wafer, and the runtime plans the token flow path. Only activations are passed between wafers, while weights remain stationary in storage. Adjacent wafers communicate over RDMA Ethernet, with single-hop latency of about 2 microseconds.
The pipelined architecture allows multiple requests to be processed simultaneously on wafers at different stages; weights do not move, and activations flow with low overhead. Under the assumption of an average request size of 256k, a 59-wafer configuration can support about 600 to 700 concurrent requests.
Traditional GPU clusters typically reach 30 to 80 tokens per second when processing a 405-billion-parameter model; WSE-3 running Llama 3.1 70B can reach about 2,100 tokens per second, while H100 SXM5 at FP8 precision is about 550 tokens per second. When concurrent users exceed 8, H100-based vLLM continuous batching can catch up with or even surpass WSE-3 through batching. Cerebras has a significant advantage in single-request, low-latency scenarios.
Kimi K3 scores 88.3% on Terminal-Bench 2.1 and 91.2% on BrowseComp; the former is slightly behind GPT-5.6 Sol's 88.8%, while the latter surpasses GPT-5.6 Sol's 90.4%. Open-source models have reached parity with top closed-source models on some agentic task benchmarks.
Agentic AI workflows are highly sensitive to latency, and multi-step tasks often require dozens of serial model calls. A speed of 1,000 to 2,000 tokens per second reduces waiting time for long-context tasks from minutes to seconds, with direct practical significance for code generation, automated research, and complex document processing.
Groq focuses on inference speed with its LPU architecture, while NVIDIA continues to advance software optimization for GPU clusters (such as vLLM and TensorRT-LLM). Cerebras's publication of a complete technical path for a 3-trillion-parameter model sets a new public reference point in the dimension of "large-model inference speed."
Whether Cerebras can commercialize the 59-wafer configuration as a stable API service or provide third-party hosted inference endpoints; whether other open-source models can reproduce similar speeds under the same hardware configuration; and whether WSE-3T wafer supply volume poses a substantive constraint on scale.
© 2026 Winzheng.com 赢政天下 | 转载请注明来源并附原文链接