Six AI Models Launch Across Voice, Robotics and Design

U.S. voice AI company Rime launched Prism on October 29, a speech-to-speech model designed to make conversations difficult to distinguish from human speech. Unlike systems that first convert speech to text and then turn a reply back into audio, Prism directly analyzes a speaker’s voice to detect speaking style, intonation and hesitation. It is designed to respond immediately when interrupted, with responses starting in about 0.57 seconds. Prism can call external tools during a conversation and can be further trained on company call data for specific work tasks. Rime says the model passed an enhanced voice Turing test and offers it through English-language APIs and SDKs for customer service and phone-work automation.

Chinese robotics AI company PokeBot unveiled RoboAPI on October 29, an API for controlling compatible robots with natural-language instructions. Its proprietary world action model, Poke-Aurora, analyzes camera footage, robot status and user instructions to predict the next action. It is designed to handle multiple tasks without collecting new task-specific data or further training the model. Developers send robot status and commands to the API, then receive model-generated action data to control the robot. PokeBot says it also tested the system at outside research facilities in the United States, China and Singapore; its technical preview API is free.

Voice AI company Hume AI launched its Audio Decision API on November 2. The model directly analyzes voices, emotions and surrounding sounds without converting speech to text, returning probabilities in response to questions. It can identify angry callers, count conversation participants, and detect sounds such as laughter and sirens. Users do not need to build a separate classification model for each question or train the system further. Answers can take the form of yes or no, multiple choice, scores, or time spans when a particular sound occurs. Hume says the model was trained on tens of millions of human evaluation examples and, in its own tests, matched the average human panel 70.7% of the time, above the 56.5% agreement rate for individual evaluators. It also reported an average score of 82% across 10 voice-evaluation benchmarks and a response latency of 0.23 seconds. Hume offers a trial environment and access requests for customer-call analysis and voice-agent quality evaluation.

Canva unveiled Brush, a vision-language model that edits designs from natural-language instructions. Brush understands images and text as well as layouts and design elements, allowing users to revise specific wording, edit multiple pages at once, change colors or mood, and turn presentations into social-media content. Canva says it preserves editable design structure and modifies the requested parts, unlike image-generation AI. The model is based on open-weight models and was further trained on real design-editing data using reinforcement learning. Canva says its internal evaluation found editing up to 69% faster and cost per task 46% lower than with existing design agents. Some features are available to Canva Pro subscribers, with wider support planned.

AI startup Nace AI released the weights of Drex Max v1, a multimodal decision model with 27 billion parameters. It processes text, images and JSON, returns probabilities for up to 255 choices in a single inference, and supports a context of up to 262,144 tokens. It scored 64 on the Artificial Analysis Intelligence Index. Nace offers an MLX-based version that runs on Apple Silicon Macs; the 4-bit quantized model works on devices with at least 24 GB of memory. Suggested uses include task selection for AI agents, request classification and policy checks. The weights are available under the open RAIL-M license.

AI optimization company Unsloth unveiled Browser Use, a feature that lets local AI models on personal computers control a web browser. Models running in the open-source Unsloth Desktop can open web pages, click buttons, fill out forms and read PDFs in response to instructions. Vision models analyze page screenshots to carry out tasks. Inference runs on the user’s computer rather than an external cloud, but internet access is required to visit external sites. Users must approve key actions, including logging in, submitting forms, making purchases and downloading files. The feature supports several open-source models with tool-calling capabilities and runs on macOS, Windows and Linux.

U.S. voice AI company Rime launched Prism on October 29, a speech-to-speech model designed to make conversations difficult to distinguish from human speech. Unlike systems that first convert speech to text and then turn a reply back into audio, Prism directly analyzes a speaker’s voice to detect speaking style, intonation and hesitation. It is designed to respond immediately when interrupted, with responses starting in about 0.57 seconds. Prism can call external tools during a conversation and can be further trained on company call data for specific work tasks. Rime says the model passed an enhanced voice Turing test and offers it through English-language APIs and SDKs for customer service and phone-work automation.

Chinese robotics AI company PokeBot unveiled RoboAPI on October 29, an API for controlling compatible robots with natural-language instructions. Its proprietary world action model, Poke-Aurora, analyzes camera footage, robot status and user instructions to predict the next action. It is designed to handle multiple tasks without collecting new task-specific data or further training the model. Developers send robot status and commands to the API, then receive model-generated action data to control the robot. PokeBot says it also tested the system at outside research facilities in the United States, China and Singapore; its technical preview API is free.

Voice AI company Hume AI launched its Audio Decision API on November 2. The model directly analyzes voices, emotions and surrounding sounds without converting speech to text, returning probabilities in response to questions. It can identify angry callers, count conversation participants, and detect sounds such as laughter and sirens. Users do not need to build a separate classification model for each question or train the system further. Answers can take the form of yes or no, multiple choice, scores, or time spans when a particular sound occurs. Hume says the model was trained on tens of millions of human evaluation examples and, in its own tests, matched the average human panel 70.7% of the time, above the 56.5% agreement rate for individual evaluators. It also reported an average score of 82% across 10 voice-evaluation benchmarks and a response latency of 0.23 seconds. Hume offers a trial environment and access requests for customer-call analysis and voice-agent quality evaluation.

Canva unveiled Brush, a vision-language model that edits designs from natural-language instructions. Brush understands images and text as well as layouts and design elements, allowing users to revise specific wording, edit multiple pages at once, change colors or mood, and turn presentations into social-media content. Canva says it preserves editable design structure and modifies the requested parts, unlike image-generation AI. The model is based on open-weight models and was further trained on real design-editing data using reinforcement learning. Canva says its internal evaluation found editing up to 69% faster and cost per task 46% lower than with existing design agents. Some features are available to Canva Pro subscribers, with wider support planned.

AI startup Nace AI released the weights of Drex Max v1, a multimodal decision model with 27 billion parameters. It processes text, images and JSON, returns probabilities for up to 255 choices in a single inference, and supports a context of up to 262,144 tokens. It scored 64 on the Artificial Analysis Intelligence Index. Nace offers an MLX-based version that runs on Apple Silicon Macs; the 4-bit quantized model works on devices with at least 24 GB of memory. Suggested uses include task selection for AI agents, request classification and policy checks. The weights are available under the open RAIL-M license.

AI optimization company Unsloth unveiled Browser Use, a feature that lets local AI models on personal computers control a web browser. Models running in the open-source Unsloth Desktop can open web pages, click buttons, fill out forms and read PDFs in response to instructions. Vision models analyze page screenshots to carry out tasks. Inference runs on the user’s computer rather than an external cloud, but internet access is required to visit external sites. Users must approve key actions, including logging in, submitting forms, making purchases and downloading files. The feature supports several open-source models with tool-calling capabilities and runs on macOS, Windows and Linux.