Kontinua combines state-of-the-art neural PDE operators (FNO, CNextU-Net, AViT) with a managed low-latency GPU cloud. Train locally on your hardware with Apache 2.0 open-source code, or scale to millisecond multi-physics inference with Kontinua Cloud.
Built on peer-reviewed research & open science from
Open-source first. Use our Python package locally with full freedom, or connect to our managed cloud for sub-50ms inference.
# 1. Install open-source package: pip install kontinua
import torch
from kontinua.core.models import CNextUNet
from kontinua.core.data import WellDataset
# 2. Load pre-trained neural surrogate from Hugging Face Hub
model = CNextUNet.from_pretrained("kontinua/cnext-unet-turbulence-2d")
model.eval()
# 3. Load initial spatiotemporal boundary state
dataset = WellDataset(well_base_path="./data", well_dataset_name="turbulence")
sample = dataset[0]["input"].unsqueeze(0)
# 4. Generate 100-step rollout in 47 milliseconds on local GPU
with torch.no_grad():
prediction = model(sample)
print(f"Output shape: {prediction.shape} | VRMSE: 0.021")# Query Kontinua Cloud Managed Triton GPU Cluster (FastAPI / gRPC)
import kontinua
client = kontinua.Client(api_key="ko_live_9f82d1...")
# Send boundary condition -> Receive validated rollout in 35ms
prediction = client.predict(
domain="compressible_euler",
mesh_input="./wing_geometry.h5",
model="cnext-unet-v2-trt", # TensorRT accelerated
rollout_steps=100,
return_metrics=True
)
print(f"Inference Latency: {prediction.latency_ms}ms | VRMSE: {prediction.vrmse}")
prediction.export_openfoam("./postProcessing/")# 1. Download 2D turbulence training split from The Well
$ kontinua download --dataset turbulence --split train --output ./data
# 2. Train Fourier Neural Operator (FNO) with Distributed Data Parallel (DDP)
$ kontinua train experiment=fno data=turbulence model.modes=32 training.batch_size=64
# 3. Benchmark local checkpoint against standardized test suite
$ kontinua benchmark --checkpoint ./runs/fno_best.pt --dataset turbulence --metrics vrmse,spectral_mse
# 4. Compile model to TensorRT engine for sub-30ms production serving
$ kontinua export --model cnext_unet --format tensorrt --fp16 --output ./engine.planSee the difference: 48 hours of traditional Navier-Stokes HPC computation vs. 47 milliseconds with Kontinua CNextU-Net surrogate.
Every architecture is fully open-sourced in PyTorch under Apache 2.0. Base weights freely downloadable from Hugging Face Hub.
Modernized U-Net + ConvNeXt Blocks
Fourier Neural Operator
Tucker-Factorized FNO
Attention Vision Transformer
Adaptive Fourier Neural Operator
Dilated Residual Network
Zero lock-in. Use the open-source library on your own infrastructure forever, or deploy on Kontinua Cloud when your team needs automated multi-GPU scale.
| Capability | Kontinua Open-Source (Community) | Kontinua Cloud (Managed Platform) |
|---|---|---|
| License & Governance | Apache 2.0 (Free Forever) | Commercial SaaS & Enterprise |
| Execution Environment | Local CPU / Workstation / Custom Cluster | Auto-Scaling H100/L4 Triton GPU Fleet |
| 8 Open Model Architectures | Full PyTorch code included | Pre-compiled TensorRT engines |
| Pre-Trained 2D Base Checkpoints | Download from Hugging Face Hub | Instant zero-setup API access |
| 3D High-Res Foundation Checkpoints | Manual training required | Included in Pro & Team tiers |
| Sub-50ms Low-Latency Inference API | Self-managed server configuration | Global REST & gRPC API with SLA |
| Managed Domain Fine-Tuning | Manual PyTorch DDP scripts | 1-Click Upload → Auto-Trained Endpoint |
| Benchmark-as-a-Service (BaaS) | Local benchmark harness | Automated CI/CD simulation tests |
| SSO/SAML, Team RBAC & SOC 2 | Not included | Enterprise Security & Compliance |
See how much time and compute spend your engineering team saves by augmenting OpenFOAM or Ansys Fluent with Kontinua surrogates.
Start for free with open source. Upgrade to Kontinua Cloud when your team needs auto-scaling GPU infrastructure.
For individual researchers and engineers running on local hardware.
For engineers integrating fast surrogate predictions into active design loops.
For simulation teams requiring managed fine-tuning & batch jobs.
Dedicated GPU clusters, on-premise air-gapped deployment, custom loss functions.
Standardized spatial (VRMSE), spectral (Fourier MSE), and temporal (Wasserstein-1) metrics across all 16 datasets.
Join our managed Cloud Early Access program. Get 3 months of Pro Cloud free upon general availability. Limited to 200 pilot seats.