Latest AI Resources

UltraEval-Audio是清华大学NLP实验室、OpenBMB和面壁智能联合开发的音频模型评测框架，最新版本为v1.1.0。专注于解决音频模型复现难、依赖冲突等问题，提供一键复现热门模型（如Vo...

Youtu-LLM 是腾讯 Youtu 团队开源的轻量级语言模型，参数规模为 19.6 亿。专为智能体任务设计，具备强大的“原生智能体能力”，在多项任务中超越同规模甚至更大模型。

RedOne is a large language model customized for social networks introduced by Little Red Book. The model is trained through a three-stage training strategy that incorporates social and cultural knowledge, strengthens multitasking capabilities, and aligns human preferences.RedOne significantly outperforms the base model in social task performance, in harmful content detection and browsing...

GLM-OCR 是智谱开源的 0.9B 轻量级专业 OCR 模型，在 OmniDocBench V1.5 以 94.6 分刷新 SOTA。兼顾“小体积”与“全场景”，扫描、手写、印章、多语混排、复杂表...

OpenReasoning-Nemotron is a series of large-scale language models open-sourced by NVIDIA to support processing of reasoning tasks in math, science and code. The models are distilled based on the DeepSeek R1 0528 model with parameter scales of 1.5B...

Ask White 5 is the flagship "All in One" model with a very high level of intelligence. The model has excellent performance in many assessments, such as the AA-Index composite assessment score of 64.7 and the STEM ability assessment score of 86, which is close to the world's leading GPT-5.

Hy3 是腾讯推出的旗舰级开源大语言模型，采用 MoE 架构，拥有 295B 总参数与 21B 激活参数，支持长达 256K 的上下文窗口。核心亮点在于快慢思考融合（Hybrid Reasoning...

Tongyi DeepResearch (Tongyi DeepResearch) is an open source intelligent body launched by Alibaba, designed for deep information retrieval and complex task reasoning, with 30 billion parameters, supporting multiple reasoning modes, including ReAct mode and deep mode...

FineVision is Hugging Face's open source visual language dataset for training advanced visual language models. It contains 17.3 million images, 24.3 million samples, 88.9 million rounds of dialog, and 9.5 billion answer tokens. The dataset aggregates...

Stand-In is a lightweight, plug-and-play identity-preserving video generation framework from Tencent's WeChat Vision team. Focusing on preserving specific identity features in video generation, it only needs to train the additional parameters of the base model 1%, and can achieve excellent results in face similarity and naturalness.

Gemini Robotics On-Device is a vision-language-action model from Google DeepMind that supports running locally in a robot. The model is able to perform tasks offline, performing fine actions based on natural language commands, such as folding clothes and pulling open bags...

Resource List Top 10 1000 Most Common Tokens 1000 Most Common Negative Tokens 20 Most Common Samplers 100 Most Common Steps 100 Most Common Dimensions 50 Most Common...

GLM-ASR is a family of high-performance speech recognition models open-sourced by Smart Spectrum AI, including the cloud-based model GLM-ASR-2512 and the open-source end-side model GLM-ASR-Nano-2512.GLM-ASR-2512 is the world's leading cloud-based speech recognition model, supporting multiple...

Ask White o4 is an innovative parallel thinking model that opens 8 thinking paths at the same time, analyzes the problem from multiple perspectives and automatically filters out the optimal solution. The model incorporates advanced Long-CoT reinforcement learning and process reward learning techniques, has powerful deep reasoning capabilities, and performs well in complex tasks.

Qwen3Guard is a fine-tuned security protection model based on the Qwen3 base model, designed for security detection. It provides accurate security categorization of prompts and responses, provides risk levels, and supports English, Chinese, and multi-language environments.Qwen3Guard comes with two pro...

Wanxing Canopy is the AIGC video creation platform launched by Wanxing Technology, covering the three major creation fields of video, picture and audio generation, which is specially designed for media and cultural industry workers, film and television/post-production workers, art and design workers, advertising and marketing practitioners, etc. to provide one-stop professional creation solutions.

DeepSeek-TUI 是美国独立开发者 Hunter Bown（GitHub: Hmbown）用 Rust 语言编写的开源终端 AI 编程智能体，专为 DeepSeek V4 系列模型（Pro/F...

Skywork Deep Research Agent v2 is a deep research intelligent body launched by Kunlun Wave, focusing on the integration and analysis of multimodal information.Skywork Deep Research Agent v2 can process text, graph...

FireRedChat is an open source full-duplex voice interaction system for Xiaohongshu with real-time bidirectional dialog capabilities and support for controlled interruptions. Adopts a modular design , including transcription control module , interaction module and dialogue manager , etc., supports cascade and semi-cascade architecture , can be flexibly deployed .

DataBuddy 是腾讯云 WeData 数据开发治理平台内置的 AI 智能体工作台，基于 WorkBuddy 同源 Agent 底层能力，采用分层 Agent 协作架构（Harness AI 理念...

Nemotron Speech ASR是英伟达开源的实时语音识别模型，专为低延迟场景优化，支持24毫秒极速转录和多人并发对话。核心采用混合Mamba-Transformer MoE架构，通过固定状态缓...

VLAC is an open source embodied reward macromodel from Shanghai Artificial Intelligence Laboratory. Based on InternVL multimodal macromodel, it integrates Internet video data and robot operation data to provide process reward and task completion estimation for robot reinforcement learning in the real world.VLAC can effectively ...

Protenix-v1是字节跳动ByteDance Seed团队推出的首个开源蛋白质结构预测模型，性能在严格对齐训练数据和模型规模后超越AlphaFold 3。模型具备显著的推理时扩展特性：通过增加采...

MobiAgent is an open source mobile intelligent body toolchain from IPADS Lab of Shanghai Jiaotong University, which helps users to build their own mobile intelligent assistants. By recording the user's operation trajectory and generating high-quality data, it trains an intelligent body that can understand natural language commands. Core features include efficient...

Magentic-UI is Microsoft's open source human-computer collaboration AI Agent research tool.Magentic-UI is based on working closely with users to facilitate complex Web tasks such as Web browsing, code execution, and file handling. The tool emphasizes collaborative planning, enabling users to raise...

Logics-Parsing is an open source Ali end-to-end document parsing model , based on Qwen2.5-VL-7B. Optimize document layout analysis and reading order inference through reinforcement learning , PDF images can be converted to structured HTML output to support a variety of content ...

Ring-2.5-1T 是蚂蚁集团百灵大模型团队开源的全球首个万亿参数混合线性架构思考模型，采用1:7 MLA与Lightning Linear Attention混合设计，激活参数量达63B。模型在...

MiMo-V2-Flash is an open source MoE architecture large model released by Xiaomi, with 309 billion total parameters and 15 billion active parameters, focusing on efficient reasoning and intelligent body applications. The model adopts hybrid attention architecture and multi-word meta-prediction technology, with an inference speed of 150 tokens/second, into...

InfinityHuman is a commercial-grade long time-series audio-driven character video generation model jointly launched by ByteDance and Zhejiang University. The model is audio-driven and can generate high-resolution, long duration and visually consistent character videos.

QVQ-Max is a state-of-the-art visual reasoning model introduced by Alitonix, which is an upgraded version of QVQ-72B-Preview. QVQ-Max is an advanced visual reasoning model that can "read" images and video content and combine the information for analysis, reasoning and problem solving.QVQ-Max's main functions include image parsing, video analysis...

Qwen3-Max-Preview is the latest flagship large language model released by Tongyi Qianwen. It is the model with the largest number of parameters in the Qwen3 family, with a parameter size of over 1 trillion. The model has significant improvements in inference, instruction following, multi-language support and long-tail knowledge coverage...

If you're looking to harness the power of artificial intelligence to assist with day-to-day tasks and automate workflows in your personal and work life, then you may be interested in the wide range of AI tools available. AssemblyAI has produced a five-minute video detailing the tools you can use to automate...

PaperBanana是北大与谷歌团队联合开源的AI学术插图自动生成框架，专门解决科研人员绘制方法示意图和统计图表的痛点。框架通过五个智能体协作（检索、规划、造型、渲染和批评），实现从文本描述到Neu...

LingBot-VA 是蚂蚁灵波开源的全球首个“自回归视频-动作世界模型”，把视频生成与机器人控制塞进同一 Transformer，每一步同时输出下一帧世界画面和对应动作，实现“边想边干”。

Meeseeks is an open source large model evaluation set used by the Meituan M17 team to evaluate the model's ability to follow instructions.Meeseeks uses a three-tiered evaluation framework to comprehensively measure whether the model is able to generate answers in strict accordance with the user's instructions from the macro to the micro level, without evaluating the knowledge of the content of the answers positively ...

HunyuanWorld-Voyager (Hunyuan Voyager for short) is the industry's first ultra-long roaming world model released by Tencent that supports native 3D reconstruction. It is a novel video diffusion framework that generates a 3D point cloud sequence of user-defined camera paths from a single image, supporting...

General Description Report mAIstro is a powerful tool designed to help users easily create customized reports through natural language processing technology. The tool utilizes LangChain technology to transform user-supplied topics and structures into detailed reports within...

Qwen3-VL-Reranker是阿里巴巴推出的多模态重排序模型，专门用于提升跨模态检索的精准度。与Qwen3-VL-Embedding协同工作：前者负责快速召回候选结果，后者通过深度跨模态交互（如...

Mixed World Model 1.5 (Tencent HY WorldPlay) is the industry's first open source real-time world modeling framework released by Tencent, covering the entire chain of data, training, and streaming inference deployment. The core is the WorldPlay autoregressive diffusion model, which uses Next-F...

Kimi K2.5 是月之暗面发布的开源旗舰模型，采用 1T MoE 架构、激活 32B、上下文 256K token，原生支持图文视频多模态输入。在 Agent、代码、视觉理解三大基准均列开源第一...

DeepSeek-OCR is an advanced optical character recognition (OCR) model open-sourced by the DeepSeek team, which converts text into images through "contextual optical compression" technology, and utilizes visual tokens for compression and decoding to achieve efficient long text processing.

Kimi Linear is a new hybrid linear attention architecture open-sourced by Dark Side of the Moon, with Kimi Delta Attention (KDA) as the core, optimizing the traditional attention model through a finer-grained gating mechanism, which significantly improves the hardware efficiency and memory control ability ...

Qwen VLo is a multimodal unified comprehension and generation model introduced by Tongyi Qianqian team. Qwen VLo can "understand" the world and recreate with high quality based on its understanding, realizing the leap from perception to generation. VLo can accurately understand the image content, and on the basis of this, it can carry out consistent and high-quality generation.

VTP (Visual Tokenizer Pre-training) is a key technology for visual generative models proposed by MiniMax Conch Video team, which enhances the performance of generative systems by improving the pre-training method of visual tokenizer (tokenizer). The traditional method...

SoulX-FlashTalk是Soul App AI团队开源的实时数字人生成模型，拥有140亿参数量，实现了0.87秒超低延迟和32帧/秒的高帧率。模型通过双向蒸馏技术解决了传统数字人延迟高、画面易...

SkyReels-V3是昆仑万维Skywork AI开源的多模态视频生成模型，被誉为视频生成领域的"全能型"标杆。模型基于"一核多支"的统一架构，在单一建模框架内集成三大核心能力：参考图像转视频、智能...

PromptEnhancer is an open source prompt word enhancement tool from Tencent's Mixed Meta team to improve the generation of text-to-image (Text-to-Image, T2I) models. Through the chain of reasoning (Chain-of-Thought, CoT) approach to the use of ...

MiniCPM 4.1 is an ultra-efficient end-side large language model introduced by Facade Intelligence. With InfLLM v2 sparse attention architecture, each lexeme only needs to calculate the relevance to less than 5% lexemes, which significantly reduces the processing overhead of long text. In a 128K long text scenario...

Egocentric-10K is a large-scale first-person view (egocentric) factory operations video dataset open-sourced by the build.ai team. The dataset contains 10,000 hours of video, totaling 1.08 billion frames, involving 2...

TurboDiffusion is a video generation acceleration framework jointly open-sourced by Tsinghua University, BioDigital Technology, and UC Berkeley, which is able to improve video generation speed by 100-200 times while maintaining nearly lossless picture quality. Through sparse linear attention, sample step distillation and 8-bit...

MindLink is a large model of open source reasoning launched by Kunlun World Wide Web. With adaptive reasoning mechanism , according to the complexity of the task can be flexibly switched inference mode , simple tasks quickly generated , complex tasks in-depth reasoning , taking into account the efficiency and accuracy . Plan-driven reasoning paradigm to remove the "think" label , down ...

Wan-Move is an open source AI video generation framework jointly developed by Ali Tongyi Labs, Tsinghua University and other organizations, focusing on high-quality video synthesis through precise motion control technology. The core technology is "potential trajectory guidance", which can seamlessly add point-level motion control to the existing image-to-video model...

GLM-4.7-Flash是智谱开源的混合专家架构语言模型，参数规模为30B，激活参数量3B，上下文窗口达200K，最大输出令牌为128K。在编程能力上表现出色，SWE-bench验证集分数达59.2...

NeuTTS Air is open source lightweight speech synthesis model, developed by Neuphonic team, which can run in real time on local devices (e.g. cell phones, laptops, Raspberry Pi) without relying on the cloud. Using 0.5B parameter Qwen architecture and self-developed NeuCodec codec...

SpatialGen is an open source 3D scene generation model of Qunar Technology, based on the diffusion model architecture, supporting the generation of spatio-temporally consistent multi-view images based on textual descriptions, reference images and 3D spatial layouts, and further generating 3D Gaussian scenes and rendering roaming videos.

Seed LiveInterpret 2.0 is a state-of-the-art simultaneous interpreting model launched by the Seed team of ByteDance, supporting two-way translation between Chinese and English. The model has near real-life translation accuracy and extremely low latency, with an average speech-to-speech delay of only 2-3 seconds, which is much lower than that of...

MonkeyCode is an open source, enterprise-grade, native AI programming assistant designed for privacy- and security-conscious development teams.MonkeyCode supports private deployment and offline use to ensure code data security. MonkeyCode supports private deployment and offline use to ensure the security of code data.

Flycount JavaAI is an intelligent Java development assistant launched by Flycount Technology. The platform supports natural language input to realize the whole process of intelligent development from requirements analysis to code generation. Developers only need to enter a description of the requirements, Flycount JavaAI can accurately understand and generate a complete engineering code framework, the platform...

ERNIE-4.5-21B-A3B-Thinking is Baidu's open source large-scale language model focused on reasoning tasks. Using the Mixed Expert (MoE) architecture , the total number of references to 21 billion , each token activates 3 billion parameters to support 128K long context window ...

SHARP (Sharp Monocular View Synthesis in Less Than a Second) is Apple's open source monocular view synthesis technology. It can quickly generate a realistic 3D representation of a scene from a single photo in less than a second...

OneCAT is a new unified multimodal model launched by Meituan in conjunction with Shanghai Jiaotong University, which adopts a pure decoder architecture and can seamlessly integrate multimodal comprehension, text-to-image generation and image editing functions. The model abandons the design of traditional multimodal models that rely on external visual coders and disambiguators through modality-specific...

AGenUI 是高德地图联合阿里千问C端应用团队发布的行业首个覆盖 iOS、Android、HarmonyOS 三端的端云一体原生 A2UI 开源框架。基于 Google 开源的 A2UI v0.9 ...

Ouro is a new type of Looped Language Models (LLMs) developed by the ByteDance Seed team, with the core innovation of directly building inference capabilities in the pre-training phase through a parameter-sharing recurrent computation structure. The model uses 24 layers as the base block through...

Wan2.2-Animate is an open source action generation model , support for action imitation and role-playing mode . Users only need to input a character picture and a reference video , the model can migrate the video character's movements and expressions to the picture character , giving the picture character dynamic expression ...

GLM-Image是智谱与华为联合开源的多模态图像生成模型，基于昇腾Atlas 800T A2芯片和昇思MindSpore框架训练，采用创新的"自回归+扩散解码器"混合架构。核心突破在于实现了国产芯片...

WebWeaver is a new dual-intelligence body framework introduced by Alibaba Tongyi team, which is mainly used in open deep research, and can simulate the human research process, which is divided into two intelligences: planning and writing.

Claude Sonnet 4.5 is an artificial intelligence model from Anthropic designed for programming, computer operations, and complex task automation. The model excels in code generation, long-duration task processing, reasoning, and mathematical computation, supporting everything from initial planning...

VoiceSculptor 是西北工业大学联合多家机构开源的音色设计模型，基于 LLaSA-3B 和 CosyVoice2 开发，专注于通过自然语言指令生成多样化音色的语音合成。支持对语速、音量、基频...

Wenxin Big Model X1.1 is a deep thinking model launched by Baidu, based on a hybrid reinforcement learning framework that focuses on improving language understanding and generation. The model excels in handling complex questions, following instructions and simulating the behavior of intelligences, and can accurately provide knowledgeable answers and high-quality text content.

Gemini 2.5 Deep Think is an AI reasoning model from Google designed to solve complex tasks. It is a variant of the model that won the gold medal at the International Mathematical Olympiad (IMO) 2025, and is designed to solve complex tasks through Parallel ...

Xiaomi-MiMo-Audio is Xiaomi's open source 7-billion-parameter end-to-end speech macromodel with powerful features such as multi-language dialog, speech continuation, less-sample generalization, and audio understanding, which is able to reach the SOTA level in speech intelligence and audio understanding benchmarks, surpassing Google Gemi...

GLM-TTS is an open source industrial-grade speech synthesis system with powerful speech synthesis capabilities. Adopting a two-stage generation architecture: the first stage will be converted to text into speech token sequences, and the second stage will be converted into high-quality audio token sequences. The system supports only 3 seconds of voice samples to complete the sound...

Youtu-GraphRAG is an open source graph retrieval augmentation generation framework from Tencent's Youtu Labs to help large language models handle complex Q&A tasks more accurately. By constructing a four-layer knowledge tree, the knowledge is disassembled into four levels of attributes, relationships, keywords and communities to realize the self-directed performance of cross-domain knowledge...

觅游（Meyo） 是美团基础研发 AI 创新产品团队推出的 AI 原生共生社区，目前已进入公测阶段。产品以"养虾"为核心隐喻，将 AI Agent 升级为拥有身份、MBTI 人格、社交关系与成长属性的...

olmOCR 2 is an open source multimodal document parsing model from the Allen Institute for Artificial Intelligence (AI2) and is an upgraded version of olmOCR. The digitized printed documents (e.g. PDF) will be high...

openPangu-VL-7B是华为开源的7B参数规模的多模态模型，专为昇腾端侧设备优化设计。模型在视觉定位、OCR识别、文档理解等任务中表现出色，支持实时推理（5FPS），单卡延迟仅160毫秒。

Yume 1.5 is an open source interactive world generation model, jointly developed by Shanghai Artificial Intelligence Laboratory, Fudan University, and Shanghai Innovation Research Institute, which is capable of real-time interactive rendering (12 FPS on a single card). It adopts the joint spatio-temporal channel modeling (TSCM) technology, even if the context length increases...

SongBloom is an open source song generation model developed by Tencent AI Lab in collaboration with The Chinese University of Hong Kong (Shenzhen) and Nanjing University, which solves the problem of "plasticity" in AI music generation, and realizes high-quality, structurally complete song generation. Simply enter 10 seconds of reference audio and corresponding lyrics, and you can...

FunctionGemma is a lightweight AI model optimized for function calls launched by Google, developed based on the Gemma 3 base model with 270 million parameters, which can convert natural language into executable API instructions in real time on cell phones, browsers and other devices. The core feature is support for local off...

StoryMem是字节跳动与南洋理工大学联合开源的AI视频生成系统，专为解决多场景视频中角色和环境一致性问题。核心通过"视觉记忆库"技术，自动存储关键帧并在后续生成时参考，确保人物外貌、服装、场景元素...

The MCP Registry is a centralized platform from GitHub that helps developers discover and install MCP servers more easily.The MCP Registry is here to help developers quickly find the AI tools they need in one place, greatly simplifying...

ZeroSearch is Alibaba Tongyi Lab open source innovative large model search engine framework. The framework does not need to interact with real search engines , based on the simulation of the search engine , with a large model of its own pre-training knowledge to generate relevant or noise documents , significantly reducing the training cost ( reduce 80% or more ...

LingBot-Depth是蚂蚁灵波科技开源的高精度空间感知模型，专门解决机器人在透明玻璃、反光物体等复杂场景中的深度识别难题。模型通过创新的"掩码深度建模"技术，在RGB图像基础上预测缺失的深度值

MOVA（MOSS-Video-and-Audio） 是上海创智学院 OpenMOSS 团队联合模思智能（MOSI）开源的端到端音视频生成模型，是中国首个高性能开源音视频模型。突破了传统"先画面后配音...

HeyGen is an AI-driven digital human video creation platform that supports a streamlined video production process, allowing users to quickly generate professional-level digital human videos. The platform is based on advanced AI technology, giving users full control over the image and voice of digital people, providing a rich library of material, including diverse background...

Granite-Docling-258M is an ultra-compact open source visual language model from IBM designed for efficient document conversion. The model converts documents into machine-readable formats while leaving layout, tables, formulas, and other elements intact.

Klear-Reasoner is a high-performance inference model from Racer, based on Qwen3-8B-Base. The model is trained by long thought chain supervised fine-tuning and reinforcement learning to perform well in mathematical and code reasoning.Klear-Reasoner...

AnyI2V is an image animation generation framework jointly launched by Fudan University, Alibaba Dharma Institute and others, which supports the conversion of static conditional images (e.g., grids, point clouds, etc.) into dynamic videos without the need for complex training processes and large amounts of data.

Clawith 是 DataElem 团队开源的"OpenClaw for Teams"——面向团队的多智能体协作平台。在保留 OpenClaw 灵魂与记忆能力的基础上，升级为 Aware 自主感知系...

VitaBench is the first interactive Agent evaluation benchmark for complex life scenarios released by the LongCat team of Meituan, assessing the comprehensive capabilities of large model intelligences in real life scenarios. The three high-frequency life scenarios of take-away ordering, restaurant dining, and traveling are used as the carrier to build the package...

Wide Research is a powerful feature of the Manus platform designed to handle complex and large-scale tasks. The platform supports hundreds of general-purpose intelligences working simultaneously through system-level parallel processing mechanisms and intelligence collaboration protocols.

灵珠是上海灵感菇智能科技有限公司打造的零门槛AI应用创作平台，由语生科学AI孵化平台推出。用户无需编写任何代码，只需在网页中输入创意想法，系统即可自动生成可实际运行的产品

EchoCare is a large model of ultrasound base developed by the Center for Artificial Intelligence and Robotics Innovation (CAIR) at the Hong Kong Institute of Innovation and Research of the Chinese Academy of Sciences (CAS), trained based on the world's largest ultrasound image dataset (more than 4.5 million images), covering multi-center, multi-region, multi-ethnicity, and more than 50 individuals...

SoulX-Podcast is Soul AI Lab's open source advanced multi-speaker conversational speech synthesis model designed for generating high quality podcast content. SoulX-Podcast has the ability to generate multiple rounds of conversations, which can simulate smooth conversations in real podcasting scenarios, and supports Mandarin, English, and multiple Chinese...

RealVideo is an open-source real-time streaming video generation system from Smart Spectrum AI that can quickly generate natural and smooth video responses in 2 to 3 seconds. Users only need to upload a photo and enter text, and the system can generate corresponding voice and video, realizing real-time conversations with AI characters...

Step3-VL-10B是阶跃星辰团队开源的100亿参数多模态AI模型，核心突破在于以轻量化设计实现顶级性能。模型通过统一预训练策略（1.2T多模态令牌数据）和创新的并行协同推理技术（PACORE...

FLM-Audio is a native full-duplex audio dialog grand model released by Beijing Zhiyuan Artificial Intelligence Research Institute in conjunction with Spin Matrix and Nanyang Technological University of Singapore, supporting both Chinese and English. Adopting native full-duplex architecture, it can merge listening, speaking and monologue at each time step...

Claude Opus 4.8 是 Anthropic 推出的最新旗舰 AI 模型，在 Opus 4.7 基础上全面升级了编码、Agentic 推理与知识工作能力，新增动态工作流（Dynamic Wo...

GELab-Zero is an open source end-side multimodal GUI Agent model by Step Leap Team , built on Qwen3-VL-4B-Instruct base model with 4B parameters.It can recognize UI elements and perform operations such as clicking and sliding, and supports cross-application tasking ...