AWS and i14 released Laplace 3B, a linear attention model, on Trainium2, saying it uses a recurrent state that remains fixed at 29.5 MB per sequence regardless of context length. It is the company's first model update since the release of the Laplace series in 2023.
AWS reported Laplace 3B trained at about 120,000 tokens per second on a trn2.48xlarge instance, measured during training. That compares with the previous model's performance on GPU-based systems, which was not specified in the source.
Laplace 3B is built on AWS Trainium2 and targets applications requiring persistent memory and predictable serving requirements, such as long-running agents and large-document analysis. Availability begins with the AWS Trainium2 platform, initially for developers and enterprises.
"We used native PyTorch support through TorchNeuron Native to bring this custom architecture to AWS Trainium2," said i14's representative. The model preserved its Gated Linear Attention (GLA) design and Muon optimizer during deployment.
The announcement follows AWS's expansion of its Trainium2 platform, which the source itself frames as a significant step in enabling efficient model training and inference for linear attention models.
AWS did not say how the model will perform on other hardware, and the source raises the open question of retrieval quality for different workloads. The engineering challenges section describes the implementation details and next steps for further optimization.
Source: awsml
