XNews

New model and multimodal tool releases

New model and multimodal tool releases

Vendors released new multimodal, audio, browser and 3D models and tools. - xAI Grok Voice Agent API launched at $0.05/min - Grok Voice: bidirectional real-time speech, multi-language, tool calls - BU-30B-A3B-Preview open model for browser control (MoE 30B) - BU-30B: 32k context, activates ~3B during run - Nano Banana Pro (Gemini image checkpoint) outperforms GPT-image for knowledge visuals - EgoX converts third-person video to realistic first-person view - Microsoft TRELLIS.2 converts images to photorealistic 3D with PBR - Claude Chrome plugin now available to paid Claude users with in-browser code testing - Mistral OCR 3: robust document parsing, supports Markdown/JSON, $2 per 1,000 pages - Seedance 1.5 Pro yields higher-quality video lip-sync and dialogue alignment

Key insights

xAI releases Grok Voice Agent API for real-time speech agents

xAI announced the Grok Voice Agent API at $0.05 per minute offering bidirectional real-time voice input/output, automatic language detection and switching, external tool calling, real-time search and reasoning, emotion-controlled voice parameters, and multiple voice options. The API is compatible with OpenAI Realtime specs and positioned as competitive on audio benchmarks.

BU-30B open model optimized for browser interaction

BU-30B-A3B-Preview is an open model tailored for browser control and DOM understanding, based on Qwen3-VL-30B-A3B-Instruct with a Mixture-of-Experts (MoE) design. It supports 32k tokens, runs with ~3B active parameters for efficiency, and aims to enable accurate web automation, clicks and extraction in single-GPU deployments.

New multimodal and 3D tools broaden content generation

Several vendors released models for visual and 3D tasks: Nano Banana Pro (Gemini image checkpoint) is cited for better knowledge visualization; EgoX reconstructs third-person video to first-person perspective; Microsoft TRELLIS.2 generates photorealistic 3D with PBR workflows; Seedance 1.5 Pro improves audio-video lip-sync and dialogue alignment for generated videos.

Mistral OCR 3 and Claude browser plugin target enterprise workflows

Mistral OCR 3 offers high-accuracy document parsing including forms, tables and handwriting, outputs Markdown/structured JSON, and is priced at $2 per 1,000 pages with batch discounts. Anthropic's Claude Chrome plugin rolled out to paid users, enabling sidebar tasks, multi-tab workflows and Claude Code integration for in-browser code testing and console reading.

Source posts