# i14 runs Laplace 3B on AWS Trainium2 with 120k tokens/sec training

> The 3‑billion‑parameter linear‑attention model trained at 120,000 tokens per second and served 11,544 tokens per second on a trn2.48xlarge instance.

Oossa · 2026-10-20 · https://oossa.com/en/i14-runs-laplace-3b-on-aws-trainium2-with-120k-tokens-sec-training

i14 used AWS Trainium2 and TorchNeuron Native to train and serve its Laplace 3B linear‑attention model. On a single trn2.48xlarge instance the model trained at roughly 120,000 tokens per second and could generate 11,544 output tokens per second across 1,280 concurrent sequences. The architecture keeps the inference state fixed at 29.5 MB per sequence, independent of context length.

## The facts

- Training speed: ~120,000 tokens/second on trn2.48xlarge
- Inference throughput: 11,544 tokens/second for 1,280 sequences

## Why it matters

Predictable memory use lets long‑running AI agents handle much larger text without growing costs.

## Sources & references

1. [Beyond the KV cache: Training and serving i14’s Laplace 3B linear attention model on AWS Trainium2 with native PyTorch](https://aws.amazon.com/blogs/machine-learning/beyond-the-kv-cache-training-and-serving-i14s-laplace-3b-linear-attention-model-on-aws-trainium2-with-native-pytorch/) – AWS, 2026-10-20

Last updated: 2026-10-20