🚀 New Features
- TensorRT-LLM engine (beta) for NVIDIA GPUs on Windows and Linux. Run local models on NVIDIA's TensorRT-LLM next to llama.cpp and MLX. Atomic Chat sets up everything the engine needs: on Linux, Docker Engine and the NVIDIA Container Toolkit; on Windows (x64 and Windows on Arm), its own WSL2 distribution with Docker inside — one administrator prompt, and Windows may ask for a restart. Requires an NVIDIA GPU of the Ampere generation or newer (RTX 30 series and up) and a driver from the R615 branch or newer.
- TensorRT-LLM models in the Model Hub. TensorRT-LLM is now a third format next to GGUF and MLX: browse curated models or search Hugging Face, see before downloading whether a model fits and runs on your GPU (and why not, if it doesn't), download it in the shared downloads panel with cancel and resume, and start a chat like with any local model. NVFP4 models need a Blackwell GPU (RTX 50 series); vision models run text-only.
🔧 Improvements & Fixes
- TensorRT-LLM setup: after a failed setup the page also offers Install… when a newer plan is available (for example after a fixed engine descriptor is published), instead of only retrying the old one.
- TensorRT-LLM troubleshooting on the provider page: a notice when the engine's sources are overridden by environment variables, Copy diagnostics (a JSON report to send to support), and Reset setup state to start the setup over. On Windows, uninstalling with "Delete app data" also removes the TensorRT-LLM WSL distribution and its data.
- Bundled atomic-core 0.9.9: WSL installation no longer hangs, an empty RebootPending state no longer blocks WSL, and on Windows on Arm the TensorRT-LLM setup downloads the arm64 engine images and TensorRT-LLM models can be checked and run (before, every model there answered that TensorRT-LLM is not available in this build).
- Keep same-origin desktop links in the app router on Windows so internal app URLs are handled in-app (while external links, downloads and fragment-only links still open normally), fixing cases where internal Windows URLs were sent to the system browser.
🙏 Contributors
Thanks to @Vect0rM, @Ooooze, @xDenside, @kisslove228, @danyurkin for their contributions to this release!