Alpha release of Infinity v3 at commit de980c3996238bfc91cb7333c835afa5a885c9b9.
Changes since alpha.2:
- Add public Mistral3 / Ministral3 text and Pixtral image support through the LibTorch backend, including text-only, image-only, and text-plus-image embedding requests (#695). Native chat templates, saved pooling configuration, and model attention settings are respected.
- Make the application environment-variable prefix configurable at build time with
ENV_VAR_BASE; the default remainsINFINITY(#694).
NVIDIA image:
docker pull ghcr.io/michaelfeil/infinity_v3:nvidia-3.0.0-alpha.3The NVIDIA image defaults to LibTorch, with Torch 2.14.1 and CUDA 13.0. It targets SM80, SM86, SM89, SM90, SM100, and SM120 with runtime kernel dispatch; SM75 is unsupported. Runtime model qualification covers H100. Other architectures require runtime qualification.
The public mistralai/Ministral-3-3B-Instruct-2512-BF16 checkpoint serves hidden-state embeddings; it is an instruction checkpoint, and this release adds no autoregressive generation endpoint or retrieval-quality guarantee. Use the NVIDIA/LibTorch image for this new model family. The separate CPU image, cpu-3.0.0-alpha.3, uses the existing Candle backend.
Ministral usage and validation · Build locally
Images are published asynchronously by the tag-triggered Infinity v3 Docker workflow. Wait for its build and image-start checks to pass before pulling the new image tags. This prerelease does not replace stable/latest image tags.
Repository, native-kernel, SDK, and Cargo dependency notices are included in the image.