Tencent, Block and IBM Lead Wave of Nine New AI Releases

Tencent has open-sourced Hy-Nomas, a world model that continues video generation according to text commands and camera movement, one of nine AI model and product releases reported by aitimes.com. The release ships model weights and code, along with a low-memory configuration that runs on a single RTX 4090.

Hy-Nomas lets users add objects or change characters' actions, background and art style mid-generation, and specify camera viewpoint and path. Block, led by Jack Dorsey, also introduced Dessa, an AI trained only on human conversational speech without text data or pretrained models. Block claims it is the first AI that learns from human voice alone to enable natural real-time conversation.

IBM released Granite Speech 5.0 2B, a 2-billion-parameter model for multilingual speech recognition and speech translation covering English, French, German, Spanish, Portuguese and Japanese. It improves recognition accuracy in noisy and reverberant environments and translation of long speech, supports up to 2000 custom keywords, and is distributed under Apache 2.0.

Ant Group's InclusionAI released the weights of Ling-3.0-flash-Sante, a medical-specialized model built on Ling-3.0-flash with 124 billion total parameters and 5.1 billion activated per token. Perplexity released Decider for Mac, an app that runs the 4-billion-parameter pplx-decider-flash-v1.2 decision model locally on Apple Silicon Macs with 16GB or more memory. It requires no API key or usage fee, sends no input to external servers, and downloads a 5.8GB model on first launch.

Baselabs under Baselten released the open-source decision-specialized models Mammut-1 (27 billion parameters) and Elephant-1 (9 billion parameters) under the MIT license. Mammut-1 scored 71.24 on the Jev Decision Index 0.4, ranking first among 208 models.

Agents, rendering and video remixing

Developer Abdulla Ababakre introduced PhoneUse, an MCP/API that lets AI agents such as Claude Code, Codex and Cursor read a phone screen and touch, type and scroll to operate mobile apps. VAST AI Research open-sourced EdenLive, a real-time generative renderer whose 5-billion-parameter model generates 720p video at over 24 frames per second on a single RTX 5090, with 5B and 14B weights released. Kylon introduced Annotations, a feature that analyzes a reference video, adds annotation-based edits, and remixes it into a new video for a team's own product.