Nine AI Models and Tools Unveiled

Tencent open-sourced Hy-Nomas on October 23, a world model that builds scenes to match text instructions and camera movements. Users can change objects, people’s actions, backgrounds, and visual styles during video generation, and specify camera views and paths. The model carries information from earlier scenes into later segments. Tencent provides the model weights, execution code, and a low-memory configuration that runs on a single RTX 4090.

Block, led by Jack Dorsey, introduced Dèṣa, a voice AI it says was trained from scratch using only human conversational audio, without text data or a pretrained model. Block says it learned real-time conversation and calls it the first AI capable of natural real-time conversation after training solely on human voice. The company says this differs from adding voice to a text-based model because Dèṣa learns conversation directly from sound.

IBM released Granite Speech 5.0 2B, a speech model with 2 billion parameters that handles English, French, German, Spanish, Portuguese, and Japanese. It keeps the previous version’s model size while improving recognition in noisy and reverberant settings and translation of long speech and conversations. Users can specify up to 2,000 keywords to help recognize product names and technical terms. It is distributed under the Apache 2.0 license. InclusionAI, part of Ant Group, released the weights for Ling 3.0 Flash Santé, a medical model based on Ling 3.0 Flash with enhanced medical knowledge, medical reasoning, literature search, and evidence synthesis. It activates 5.1 billion of its 124 billion total parameters per token and processes up to 256,000 tokens. It supports tool use and collaboration among multiple agents, but its developers said it is for research, not an independent diagnostic or prescribing medical device.

Perplexity introduced Decider for Mac, which runs its 4-billion-parameter decision model, pplx-decider-flash-v1.2, on Mac computers. The model reads text and images and returns probabilities for classification or selection, with possible uses including routing customer inquiries and sorting documents or screenshots. Input stays off external servers, and no API key or usage fee is required. It requires an Apple silicon Mac with at least 16 GB of memory; the 5.8 GB model downloads on first launch. Baseten subsidiary Basetlabs released Mammoth-1, with 27 billion parameters, and Elephant-1, with 9 billion. Both produce probabilities for different options based on context and questions, and are available under the MIT license. The developers said Mammoth-1 ranked first among 208 models on the Jev Decision Index 0.4, scoring 71.24. Potential uses include sorting customer inquiries, assigning AI tasks, and safety checks.

Developer Abdullah Ababakr unveiled Phoneuse, which lets AI agents read smartphone screens and control apps through taps, typing, and scrolling. It can connect to Claude Code, Codex, and Cursor through MCP or an API, and offers cloud-based Android devices as well as physical smartphones. For physical iPhones, it can also read the screen with a camera and operate the device with a physical stylus.

VAST AI Research open-sourced EdenLive, which turns game screens or reference videos into AI-generated video in real time. Given a game engine’s screen, a desired first frame, and text instructions, it renders scenes in a photorealistic or other visual style while following object placement and movement. The developers said its 5-billion-parameter model generates 720p video at more than 24 frames per second on a single RTX 5090. Existing software manages game rules and object positions while AI handles the screen’s appearance. The team provides weights for 5-billion- and 14-billion-parameter models. Kylone introduced Annotation, which analyzes reference videos and lets users add notes describing changes to create new videos tailored to its product. The collaboration feature lets team members combine feedback on directing and preferences; it is not the release of weights for a new generative model.

Tencent open-sourced Hy-Nomas on October 23, a world model that builds scenes to match text instructions and camera movements. Users can change objects, people’s actions, backgrounds, and visual styles during video generation, and specify camera views and paths. The model carries information from earlier scenes into later segments. Tencent provides the model weights, execution code, and a low-memory configuration that runs on a single RTX 4090.

Block, led by Jack Dorsey, introduced Dèṣa, a voice AI it says was trained from scratch using only human conversational audio, without text data or a pretrained model. Block says it learned real-time conversation and calls it the first AI capable of natural real-time conversation after training solely on human voice. The company says this differs from adding voice to a text-based model because Dèṣa learns conversation directly from sound.

IBM released Granite Speech 5.0 2B, a speech model with 2 billion parameters that handles English, French, German, Spanish, Portuguese, and Japanese. It keeps the previous version’s model size while improving recognition in noisy and reverberant settings and translation of long speech and conversations. Users can specify up to 2,000 keywords to help recognize product names and technical terms. It is distributed under the Apache 2.0 license. InclusionAI, part of Ant Group, released the weights for Ling 3.0 Flash Santé, a medical model based on Ling 3.0 Flash with enhanced medical knowledge, medical reasoning, literature search, and evidence synthesis. It activates 5.1 billion of its 124 billion total parameters per token and processes up to 256,000 tokens. It supports tool use and collaboration among multiple agents, but its developers said it is for research, not an independent diagnostic or prescribing medical device.

Perplexity introduced Decider for Mac, which runs its 4-billion-parameter decision model, pplx-decider-flash-v1.2, on Mac computers. The model reads text and images and returns probabilities for classification or selection, with possible uses including routing customer inquiries and sorting documents or screenshots. Input stays off external servers, and no API key or usage fee is required. It requires an Apple silicon Mac with at least 16 GB of memory; the 5.8 GB model downloads on first launch. Baseten subsidiary Basetlabs released Mammoth-1, with 27 billion parameters, and Elephant-1, with 9 billion. Both produce probabilities for different options based on context and questions, and are available under the MIT license. The developers said Mammoth-1 ranked first among 208 models on the Jev Decision Index 0.4, scoring 71.24. Potential uses include sorting customer inquiries, assigning AI tasks, and safety checks.

Developer Abdullah Ababakr unveiled Phoneuse, which lets AI agents read smartphone screens and control apps through taps, typing, and scrolling. It can connect to Claude Code, Codex, and Cursor through MCP or an API, and offers cloud-based Android devices as well as physical smartphones. For physical iPhones, it can also read the screen with a camera and operate the device with a physical stylus.

VAST AI Research open-sourced EdenLive, which turns game screens or reference videos into AI-generated video in real time. Given a game engine’s screen, a desired first frame, and text instructions, it renders scenes in a photorealistic or other visual style while following object placement and movement. The developers said its 5-billion-parameter model generates 720p video at more than 24 frames per second on a single RTX 5090. Existing software manages game rules and object positions while AI handles the screen’s appearance. The team provides weights for 5-billion- and 14-billion-parameter models. Kylone introduced Annotation, which analyzes reference videos and lets users add notes describing changes to create new videos tailored to its product. The collaboration feature lets team members combine feedback on directing and preferences; it is not the release of weights for a new generative model.