The latest release label for zhongkaifu/tensorsharp is v2026.10.23. The repository is a .NET 10-based inference engine and agent runtime that runs GGUF models locally. It provides a console app, a web chat interface, an iPhone app, and an Ollama/OpenAI-compatible HTTP API. However, the given material lists nothing new for this version, so the specific release changes cannot be confirmed.
Beyond running models, the project supports text and reasoning, multimodal processing, embeddings, image and video generation, Agent Skills, and code tools. It has a C# CPU backend plus CUDA, Metal, and Vulkan backends, and the TensorAgent app can be used on iPhone, iPad, Mac, and Windows. Sparse attention is a method that skips some attention computation. The research summary says a dedicated CUDA kernel is 9.8 times faster at 64K prefill, but it gives no comparison baseline. The repository uses the BSD-3-Clause license, and feature and backend support vary by model.
So the breadth of supported features and the specific performance claims should be read separately. The README summary describes several interfaces and task types, but it does not say that every feature works on every model. The 9.8 times at 64K prefill likewise comes without a stated baseline, so it cannot be interpreted as a performance comparison with other engines. The change list for the latest release is also missing from the material, so there is no ground for judging what was added or fixed in v2026.10.23.