How we deploy Trillion-parameter frontier models on Cerebras

Oct 22 2026 — Cerebras explains how it splits trillion-parameter frontier models across multiple WSE-3T wafers (44 GB on-chip memory each) to run multi-trillion-parameter models at 1000–2000+ tokens per second.

Uses open-source Kimi K3 (~3T parameters, 69 linear-attention layers, 24 dense layers) as an illustrative example. Kimi K3 scores 88.3% on Terminal-Bench 2.1 (vs GPT-5.6 Sol 88.8%) and 91.2% on BrowseComp (vs GPT-5.6 Sol 90.4%).

Example assumptions: 256k average request size, 4-bit weight quantization, 8-bit KV cache. With 3T parameters at 4 bits, ~16 GB per layer across 93 layers. Dense layers: one per wafer plus KV cache SRAM; linear-attention layers: two per wafer (no per-token KV). Totals: 35 wafers for linear layers + 24 for dense = 59 wafers.

Pipeline execution: compiler maps layers; runtime routes tokens via RDMA over Ethernet (~2 µs per hop). Weights stay on-chip; only activations cross wafers. Concurrent requests: pipeline kept full with micro-batches; KV cache on dense wafers caps concurrency (example math yields thousands of concurrent 256k-context requests).

Same approach scales beyond 3T: add wafers for weights and request state. Post ties to prior OpenAI Ultrafast / Cerebras inference partnerships.