Overview
HyPDE is a 323-million-parameter, fused-history transformer foundation model for forecasting two-dimensional partial differential equation (PDE) simulations. Tencent says it was developed with Hy4 and the Hyra 2.0 research framework. (github.com)
Capabilities and uses
Given the last six frames of a simulated physical field on a regular 2D grid, HyPDE predicts the next four frames. It can feed predictions back into the model for longer rollouts. A single set of weights covers 13 physical systems, including acoustics, compressible flow, reaction-diffusion and shallow-water simulations. It is intended for scientific simulation and physical-field forecasting. (raw.githubusercontent.com)
In Tencent's zero-shot evaluation on test splits from 13 The Well and FlowBench datasets, HyPDE had lower median variance-normalized RMSE than the released Walrus model on 9 of 13 datasets for one-step forecasts, 10 of 13 for rollout frames 1–20, and 8 of 13 for frames 21–60. Neither model was fine-tuned for the comparison. (raw.githubusercontent.com)
Availability and limitations
The code, Python package and approximately 1.3 GB FP32 weights are available for self-hosted inference under Apache-2.0. Python 3.10+ and PyTorch 2.4+ are required; a CUDA GPU is recommended, while CPU inference is slower. (raw.githubusercontent.com)
The published model is for regular 2D grids. Tencent notes that other systems using fields in its vocabulary can be supplied, but those systems have not been tested. Forecast accuracy varies by dataset and rollout horizon; the reported results do not establish superiority on every dataset. The converted FlowBench evaluation files are not included in the release. (raw.githubusercontent.com)
Pricing
As of November 14, 2026, HyPDE-specific hosted API pricing, including input and output charges, could not be verified in either the Tencent Cloud China pricing page or the Tencent Cloud International pricing page. The published self-hosted weights and code do not establish a hosted API price; users may incur their own computing costs. (cloud.tencent.com)
Release and official sources
- Release date: November 4, 2026, according to the official Hyra Results announcement. (github.com)
- Official HyPDE documentation and model files. (github.com)