Klarna has launched a voice AI customer service agent built on ElevenLabs' ElevenAgents in the US. Serving as the first point of contact for phone support, it targets all 35 million US users and achieves up to a 10x improvement in resolution speed for handleable queries. The agent supports natural conversation, context understanding, and seamless handoff to human agents, with plans to expand to 114 million users globally.
ElevenLabs officially announced the launch of ElevenLabs for Government, a suite of voice and chat agents designed for the public sector with government-grade security and compliance. It supports multilingual, 24/7 natural conversations. Deployments include Ukraine (voice layer for digital humans), the Czech Republic (handling approximately 5,000 hotline calls per day with an independent resolution rate of about 85%), and Midland City, USA (replacing traditional IVR, expected to reduce missed calls by approximately 7,000 per month).
Google Research has released DialogLab, an open-source prototype framework presented at ACM UIST 2025 that supports the creation, simulation, and validation of dynamic human-machine group conversations. The framework decouples conversations into group dynamics (participants, roles, subgroups) and conversational flow dynamics (segments, turns, interaction styles). It provides a visual drag-and-drop interface, human-AI collaborative simulation modes, and diagnostic validation dashboards. Its usability in education, game design, and social science research was validated through evaluations by 14 domain experts.
The Tongyi Qwen team has released Qwen-Image-2.0, a new-generation image generation foundation model that supports 1k token instructions and 2k resolution (2048×2048), featuring integrated text-to-image and image-to-image capabilities. It emphasizes accuracy, complexity, aesthetic layout, material realism, and alignment consistency in text rendering, achieving superior performance in both text-to-image and image-to-image benchmarks in AI Arena blind tests. The model architecture is more lightweight, enabling faster inference.
Google Research and Google DeepMind published a paper at the NeurIPS 2025 workshop introducing Perch 2.0, a bioacoustic foundation model primarily trained on audio from terrestrial vocalizing animals such as birds; despite containing no underwater audio in its training data, it performs exceptionally well on marine acoustic tasks like whale call classification.
Google DeepMind published two papers introducing the professional research applications of the Gemini Deep Think mode in mathematics, physics, and computer science: its mathematical research agent Aletheia can iteratively generate.
Moonshot AI has released Kimi Agent Swarm, a self-organizing agent architecture that supports up to 100 sub-agents running in parallel, over 1,500 tool calls per task, and an inference speed 4.5 times faster than serial mode; this feature is now available to top-tier subscription users and is suitable for scenarios such as large-scale information discovery, generation of hundred-page literature reviews, and multi-perspective collaborative analysis.
Moonshot AI has released Kimi Agent Swarm, supporting up to 100 parallel sub-agents for Kimi K2.5, executing over 1,500 tool calls, with inference speed reaching 4.5 times that of serial mode; this feature is now available to top-tier subscribers and is suitable for scenarios such as large-scale discovery, batch document processing, and multi-perspective analysis.
Hugging Face officially announced the release of Transformers.js v4. Core upgrades include a new WebGPU runtime written in C++, supporting various JavaScript environments such as browsers, Node, Bun, and Deno.
ServiceNow has released SyGra 2.0.0, introducing a new Studio interactive environment that supports visual orchestration of synthetic data generation workflows on a canvas. It integrates model configurations (OpenAI, Azure OpenAI, Ollama, Vertex, Bedrock, vLLM, etc.), multi-source data connections (Hugging Face, file systems, ServiceNow), inline variable prompts, real-time execution flow monitoring, node-level debugging, and execution history.
OpenAI has launched Frontier, an enterprise AI agent platform that supports building, deploying, and managing AI agents with shared context, onboarding workflows, permission controls, and governance capabilities.
Mistral releases Voxtral Transcribe 2, which includes Voxtral Mini Transcribe V2 for batch transcription and Voxtral Realtime for real-time applications.
OpenAI has released the Codex App Server, a bidirectional JSON-RPC API that supports streaming progress, tool calls, human approval, and diff comparison for embedding Codex agents.
Hugging Face has launched a Community Evals feature on the Hub, supporting the registration of dataset repositories as benchmarks (such as MMLU-Pro, GPQA, and HLE). Model repositories store scores via .eval_results/*.yaml files. Community users can submit evaluation results through pull requests (PRs) with source links attached. All results are publicly traceable and support aggregation via the Hub API. The feature is currently in beta.
H Company released Holo2-235B-A22B Preview, its largest UI localization model to date. It achieves 78.5% on Screenspot-Pro (agent mode, 3 steps) and 79.0% on OSWorld G, supporting agentic iterative localization for high-resolution 4K interfaces. The model has been open-sourced on Hugging Face, with training using SkyPilot to uniformly schedule multi-cloud Kubernetes clusters.
Hugging Face published the second part of its PRX text-to-image model training design series on its blog, systematically evaluating representation alignment (REPA/iREPA/REPA-E) and training objectives (contrastive flow matching/JiT).
On February 2, Baidu Intelligent Cloud released the Red Finger Operator mobile access solution, enabling the deployment of OpenClaw on mobile devices through its ARM cloud phone service. It supports a bidirectional mode where OpenClaw acts as the master controller invoking Operator for mobile app interactions, and an upcoming mode where Operator acts as the master controller delegating tasks to OpenClaw sub-modules. Standardized APIs and MCP servers are provided; the web version and Android APP have been launched, applied in marketing automation for industries such as automotive and consumer goods.
The Moonshot AI team released the WorldVQA evaluation benchmark to assess the factual accuracy of multimodal large language models (MLLMs) regarding atomic-level visual world knowledge. The dataset contains 3,500 high-quality image-text question-answer pairs, covering 9 categories of topics and explicitly distinguishing between head and long-tail knowledge. All samples underwent multi-stage human verification. Experiments show that state-of-the-art models often have accuracy below 50% on long-tail knowledge and generally exhibit severe overconfidence. The benchmark has been open-sourced, including links to the paper, code, and dataset.
The Moonshot AI team released the WorldVQA benchmark to evaluate the factual accuracy of Multimodal Large Language Models (MLLMs) concerning atomic-level visual world knowledge. The benchmark contains 3,500 high-quality image-question pairs covering 9 categories, with a clear distinction between head and long-tail knowledge. Experiments show that frontier models often achieve accuracy below 50% on long-tail knowledge. The team simultaneously open-sourced the dataset, evaluation scripts, paper (arXiv:2602.02537), code (GitHub), and data (Hugging Face).
Tongyi has launched the open-source model Qwen3-Coder-Next, based on Qwen3-Next-80B-A3B-Base. It employs a hybrid attention and MoE architecture, trained through executable task synthesis, environment interaction, and reinforcement learning. Using the SWE-Agent framework, it achieves over 70% on SWE-Bench Verified, with only 3B active parameters matching models with 10×–20× more active parameters.
Zhipu officially released and open-sourced the GLM-OCR model, with only 0.9B parameters, achieving SOTA on OmniDocBench V1.5 with a score of 94.6; it supports recognition in high-difficulty scenarios such as handwriting, complex tables, seals, and code documents, outputting structured JSON and HTML; GitHub and Hugging Face open-source addresses are provided, supporting vLLM/SGLang/Ollama deployment, with API pricing at 0.2 yuan/million Tokens.
ElevenLabs officially announced that the Eleven v3 text-to-speech model has exited the Alpha stage and is now fully available for use. Compared to the Alpha version, user preference reached 72%, and the overall error rate for symbol and number interpretation decreased by 68%, dropping from 15.3% to 4.9%. The model covers 8 languages and 27 types of scenarios, including context-sensitive use cases such as currency, chemical formulas, and sports scores.
OpenAI has released the Codex macOS app, serving as a command center for AI coding and software development, supporting multi-agent systems, parallel workflows, and long-running tasks.