New AI models and tools

The Tencent Hy Team released an open-weight mixture-of-experts text model on August 28, 2026, aimed at coding, office analysis, finance, game development and scientific research. It is claimed to have 770 billion total and 49 billion active parameters and a one-million-token context, with reasoning and no-think modes, vLLM and SGLang support, and full and FP8 weights. The limitations: it is an early preview, with long thinking and heavy self-verification on complex tasks, and hardware demand and benchmark reproducibility are concerns. Weights are downloadable under the repository licence; no official hosted-API pricing was found.

Ukraine's defence digital ecosystem released an in-product knowledge and support assistant on August 28, with system-wide availability for authorised DELTA users. It answers documentation questions and routes unresolved issues to responsible teams. It is not public, and its security architecture, underlying model and independent accuracy evaluation have not been released.

A model-agnostic hardware interface specification, developed initially with HHMI Janelia, was released on August 27 as a limited research preview, with an open-source release planned. Pricing has not been announced, and it remains an early-stage effort without established certification, liability or cross-vendor safety assurance.