Cerebras Explains Trillion-Parameter Model Deployment
Original: Cerebras Deploys Trillion-Parameter AI Models
Why This Matters
Shows a viable path to serving trillion-parameter open models at commercial inference speeds.
Cerebras published a technical breakdown of how it deploys multi-trillion-parameter models across WSE-3T wafers, using Kimi K3's ~3T parameters as the example, achieving 1,000–2,000+ tokens per second.
Cerebras detailed its approach to running frontier-scale models across clusters of WSE-3T wafer-scale chips, each carrying 44 GB of on-chip SRAM. Using Kimi K3—an open-source 3-trillion-parameter model with 69 linear-attention layers and 24 dense layers—as a worked example, the post walks through how the company fits weights into memory and pipelines inference.
With 4-bit quantization, Kimi K3's full weight payload is roughly 1.5 TB. Each of its 93 layers occupies about 16 GB. Dense layers, which require KV cache storage proportional to active requests, get one layer per wafer. Linear-attention layers carry a fixed-size state regardless of context length, so two fit per wafer. That arithmetic yields 59 wafers total for the full model.
Weights stay fixed in place on each wafer. Only activations move between wafers via RDMA over Ethernet, with roughly 2-microsecond hops. Cerebras frames the setup as an assembly line: each wafer is a pipeline stage, and tokens travel through sequentially. The company claims this architecture supports 1,000–2,000+ tokens per second at scale.
Kimi K3 itself scores 88.3% on Terminal-Bench 2.1 and 91.2% on BrowseComp, edging GPT-5.6 Sol on the latter benchmark.