InstinctFlash Enables Real-Time 5B World-Action Models on Jetson Thor and RTX GPUs

Full-Source Release for Robotics Models

InstinctFlash has announced the full-source release of its platform, which includes eight robotics model families, acceleration kernels, and Python/WebSocket serving capabilities through a single Runtime. This release aims to facilitate the deployment of large world-action models in real-time robotics applications.

Performance Benchmarks on Jetson Thor

Benchmarks conducted on NVIDIA Jetson Thor show significant performance improvements. InstinctFlash achieved up to a 33.78x speedup with the LingBot-VA model, utilizing FP8 quantization and fewer sampling steps. This optimization was observed without a loss in task performance during real-robot tests, indicating efficient inference for complex models.

Expanded GPU Support

In addition to Jetson Thor, InstinctFlash now supports NVIDIA RTX 4090 and RTX 5090 GPUs. This expansion allows developers to deploy and run these 5B world-action models on high-end workstations using the same Runtime API, providing flexibility for development and deployment environments.

Simplified Deployment and Inference

The platform offers a simplified deployment process. Users can point the 'serve' command at a fine-tuned checkpoint, and InstinctFlash automatically detects the model family, generates a declaration, and starts serving. For stock releases, models can be loaded using their Hub ID after installing the family's environment, streamlining the inference workflow.