Latest AI Resources

ViMax是香港大学数据科学实验室开源的多智能体视频生成框架，能实现从创意输入到视频输出的全流程自动化。整合了剧本生成、分镜设计、镜头规划和视频渲染等功能，支持用户通过自然语言描述生成连贯的影视级视频...

FLUX.2是Black Forest Labs发布的开源图像生成与编辑模型，支持文生图、多图参考和图像编辑，具备更丰富的细节、清晰纹理和稳定光线。分为四个版本：FLUX.2 [pro]（媲美顶级闭源...

Fara-7B是微软开源发布的70亿参数规模的计算机操作代理（CUA）模型，基于Qwen2.5-VL-7B架构。通过视觉解析网页截图，在屏幕上执行点击、输入等操作，无需依赖额外的可访问性树或多个大模型...

HunyuanOCR是腾讯混元团队开源的高性能光学字符识别模型，参数量仅10亿。基于混元多模态架构开发，采用端到端设计，能高效处理文字检测、识别及文档解析任务。模型在复杂文档测试中得分94.1分，超越...

Supertonic是开源的高性能的文本转语音（TTS）系统，专注于在本地设备上快速生成语音。采用ONNX Runtime技术，可在手机、电脑甚至树莓派等设备上运行，支持23种语言和语音克隆，无需网络...

MiMo-Embodied是小米集团开源的全球首个成功融合具身智能（Embodied AI）与自动驾驶的跨具身基础模型。解决具身智能与自动驾驶之间的知识迁移难题，实现两大领域的任务统一建模。

MOSS-Speech是复旦大学邱锡鹏教授团队开源的语音到语音（Speech-to-Speech）大模型。突破传统语音处理方式，无需文本引导，直接对语音进行理解和生成，能捕捉语调、情绪等非文字要素，使...

Parallax是分布式AI实验室Gradient开源的全球首个“全自主AI操作系统”。支持在Mac、Windows等异构设备上跨平台部署大模型，让用户完全掌控模型、数据与AI记忆。系统内置网络感知分...

HunyuanVideo 1.5 是腾讯混元大模型团队开源的轻量级视频生成模型，基于 Diffusion Transformer（DiT）架构，参数量为 8.3B。支持生成 5-10 秒的高清视频，分...

Awex是蚂蚁集团开源的高性能权重交换框架，专为强化学习中的大规模参数同步设计。能在秒级完成TB级参数交换，显著提升训练推理效率。Awex具备极速同步性能，在千卡集群上，万亿参数模型可在6秒内完成全量...

Seekdb（OceanBase Seekdb）是蚂蚁OceanBase开源的 AI 原生混合搜索数据库，支持向量、全文、标量及地理空间数据的统一混合搜索，采用多阶段检索机制，实现低延迟下的高精度搜索...

LoopTool是上海交通大学与小红书团队开源的自动化的工具调用数据进化框架，专为提升大语言模型的工具调用能力设计。通过闭环迭代优化数据生成与模型训练，利用开源模型（如Qwen3-32B）作为数据生成...

SAM 3D是Meta公司推出的基于SAM系列的3D重建模型，包含SAM 3D Objects和SAM 3D Body两个分支。其中SAM 3D Objects能从单张照片生成可交互的3D物体模型，支...

AgentEvolver是阿里巴巴通义实验室开源的智能体进化系统。通过自我提问、自我导航和自我归因三种机制，实现智能体的自主学习与进化。AgentEvolver采用服务导向架构，将环境沙盒、LLM和经...

MemOS是MemTensor等开源的面向大语言模型（LLM）的记忆管理与调度框架。将记忆视为与算力同等重要的资源，通过标准化的MemCube记忆单元，统一管理明文、激活状态和参数记忆。

WithAnyone 是复旦大学和阶跃星辰联合开发的 AI 合照生成模型，解决传统 AI 图像生成中常见的“复制粘贴”问题，实现更自然、可控的多人图像生成。模型通过大规模数据集 MultiID-2M ...

ChatTutor是开源的AI教学辅助工具，专注于STEM学科的可视化互动学习。通过多智能体架构实现对话式答疑和动态绘图功能，能在电子白板上实时绘制数学图形、物理电路或思维导图，帮助用户直观理解抽象概...

DPAI Arena（Developer Productivity AI Arena）是JetBrains创建的开放基准测试平台，衡量AI辅助开发工具在真实世界软件工程任务中的有效性。通过透明的评估流...

EverMemOS是陈天桥领导的盛大团队推出的开源长期记忆操作系统，专为AI智能体设计，解决大语言模型因固定上下文窗口导致的记忆断裂问题。系统基于人类大脑记忆机制，采用四层架构（代理层、记忆层、索引层...

Astron Agent是科大讯飞开源的企业级智能工作流开发平台，专注于帮助企业快速构建可落地的AI代理应用。采用Java+Spring Boot技术栈，支持轻量化私有化部署（最低2核4G配置），内置...

Bee是腾讯混元团队与清华大学联合推出的全栈开源多模态大模型解决方案，通过提升数据质量缩小开源模型与闭源模型的性能差距。项目包含三大核心成果：1500万规模的高质量双层CoT数据集Honey-Data...

InfinityStar是字节跳动开源的统一时空自回归框架，专为高分辨率图像和视频生成设计。采用离散自回归方法，能在单一模型中同时处理文本到图像、文本到视频、图像到视频等任务。框架在VBench基准测...

Koina是开源的去中心化机器学习平台，专注于简化蛋白质组学数据分析。由德国慕尼黑工业大学和美国密歇根大学团队开发。平台通过标准化接口整合了30多个主流模型（如ProSIT、MS²PIP），支持肽段质...

VibeThinker-1.5B是微博AI开源的15亿参数的大型语言模型。基于阿里巴巴的Qwen2.5-Math-1.5B进行精细调整，专为数学和代码任务优化，表现出色，推理性能在行业内处于领先水平。

BestBlogs 是专注于为技术从业者、创业者和产品经理等提供高质量内容的平台。通过 RSS 订阅和爬虫技术，从 400 多个优质博客中收集文章、播客、视频等多形式内容。核心优势在于利用 AI 大语...

Egocentric-10K 是大规模的第一人称视角（egocentric）工厂操作视频数据集，由 build.ai 团队开源。数据集包含 10,000 小时的视频，总帧数达 10.8 亿帧，涉及 2...

LazyCraft 是商汤基于开源框架 LazyLLM 构建的开源 AI Agent 应用开发与管理平台，为企业和开发者提供一站式AI应用开发解决方案。帮助开发者以低门槛、低成本快速构建和发布大模型应...

Kosong 是月之暗面（Moonshot AI）开源的全新AI Agent开发框架，为开发者提供一个轻量、灵活且高度可扩展的底层支持，以构建下一代智能体应用。通过异步工具编排引擎，能高效调度多个工具...

SenseNova-SI是商汤科技发布的开源空间智能大模型，专注于提升AI在空间理解与推理方面的能力。模型在空间测量、重构、关系判断、视角转换、形变分析和空间推理等六个核心维度上表现出色，显著优于其他...

Omnilingual ASR是Meta推出的多语语音识别框架，覆盖1600+语言，78%语言字符错误率低于10%。其70亿参数wav2vec 2.0编码器结合CTC与Transformer解码器，支...

Frappe Builder是开源的低代码建站工具，由Frappe公司开发，核心特点是提供类似Figma的可视化编辑器，支持拖拽组件快速搭建网站。属于Frappe生态（Frappeverse）的一部分...

DeepOCR 是开源复刻项目，实现 DeepSeek-OCR 的核心架构，通过光学压缩技术高效处理文本信息。核心是 DeepEncoder，由 SAM-base（处理高分辨率图像）、16×卷积压缩器...

Glow 是开源的命令行工具，用于在终端中优雅地渲染 Markdown 文件。工具支持高亮代码块、数学公式等复杂元素，提供丰富的功能，如自定义样式、分页显示、鼠标支持等。

NocoBase是基于AI驱动的开源无代码开发平台，支持快速搭建业务系统，无需编程即可通过配置完成应用开发。项目采用Apache-2.0协议，提供私有化部署和灵活扩展能力，适用于企业管理、协作平台等场...

UniWorld V2是兔展智能与北京大学UniWorld团队联合推出的新一代图像编辑模型。在图像编辑领域具有显著优势，特别是在中文理解和复杂指令执行方面表现出色。模型能精准渲染艺术中文字体，支持精细...

SmartResume 是阿里巴巴开源的智能简历解析与优化工具，能高效地从 PDF、图片或 Office 文档中提取结构化信息，如基本资料、教育经历和工作经验等。通过融合 OCR 技术和 PDF 元数...

Step-Audio-EditX是开源的音频编辑大模型，由阶跃星辰团队研发，专注于通过人工智能技术实现音频内容的精细操控。模型能动态调整音频的情绪、说话风格（如撒娇、老人腔等）和副语言元素（如笑声、叹...

Open-o3 Video 是北京大学和字节跳动联合开发的开源视频推理模型，专注于通过时间和空间证据增强视频推理能力。通过明确标注关键证据的时间戳和边界框，帮助模型更好地理解和解释视频内容。

Handy is open source and free local speech to text tool, supporting Windows, MacOS and Linux systems, developed by Rust and React. It is suitable for quick transcription and text input by processing voice data locally without uploading it to the cloud to ensure privacy and security.

FG-CLIP 2 is the world's leading graphical cross-modal visual language model (VL-M) launched by 360 Artificial Intelligence Research Institute, which surpasses similar models from Google and Meta in 29 authoritative benchmark tests, making it the most powerful VL-M at present.It is able to accurately recognize the gross...

BettaFish is an open source multi-intelligence system for public opinion analysis. Using multi-intelligent body architecture, through Query, Media, Insight, Report and other Agents work together to achieve retrieval, extraction and reporting closed loop. The system supports AI-driven full ...

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...

ChronoEdit, an open-source AI image editing framework developed by NVIDIA in conjunction with the University of Toronto, redefines the image editing task as a video generation task to ensure that the editing results are temporally and physically consistent. By distilling a pre-trained video generation model with 14B parameters from a...

LongCat-Flash-Omni is an open source fully modal big language model released by the LongCat team of Meituan. With a parameter scale of 560 billion (27 billion activated parameters), it realizes millisecond-level real-time audio and video interaction capabilities while maintaining a large number of parameters.

Petri is an open source AI security auditing framework developed by Anthropic that systematically assesses the security and behavioral alignment of AI models. By simulating a real-world scenario where an automated auditor engages in multiple rounds of conversations with a target model, followed by a judge agent that acts on the model's...

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 ...

FIBO is the world's first open source text generation image model with native JSON support developed by Bria AI. Based on the DiT (Diffusion Transformer) architecture with 8B parameters, it adopts the Flow Matching training method...

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...

GigaBrain-0 is the first end-to-end Vision-Language-Action (VLA) embodied base model in China that uses world model generation data to realize real machine generalization, and it is jointly released as open source by GigaVision and Hubei Humanoid Robot Innovation Center. It adopts the hybrid Transformer architecture, integrating ...

Ming-flash-omni-Preview is an open-source full-modal macromodel released by Ant Group inclusionAI, with a parameter scale of hundreds of billions, based on the sparse MoE architecture of Ling 2.0, with total parameters of 103B and activations of 9B. in full-modal understanding and generating...

OmniVinci is an open-source, fully modal large-scale language model developed by NVIDIA that solves the problem of modal fragmentation in multimodal models through architectural innovation and data optimization. Alignment of visual and audio embeddings is enhanced by OmniAlignNet, which utilizes temporally embedded group capture...

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...

ValueCell is an open source multi-intelligent body financial application platform that improves the efficiency of financial analysis and investment management through AI technology. Simulating a professional investment team, multiple AI intelligences work together, covering market analysis, sentiment analysis, fundamental research, automated trading and other functions, to provide users with a comprehensive...

Dexbotic is the open source Visual-Linguistic-Action (VLA) model of embodied intelligence one-stop scientific research service platform of Dexmal, which solves the problems of fragmentation and low efficiency of research in the field of embodied intelligence. Based on PyTorch, Dexbotic is a one-stop research service platform to solve the problems of fragmentation and inefficiency in the field of embodied intelligence...

LongCat-Video is a 1.36 billion parameter video generation model open source by the LongCat team, using the MIT open source protocol, supporting three major tasks: text-generated video, graph-generated video and video continuation. The model through the "coarse to fine" generation strategy and block sparse attention mechanism, can be in a number of minutes ...

DreamOmni2 is a multimodal AI image editing and generation model open-sourced by Jiajia's team at HKUST. Can handle both text and image commands, supports multiple reference images, providing creators with more flexible ways of creation. The model is trained using a three-stage data synthesis process , joint training generation/editing...

WorldMirror 1.1 (WorldMirror) is an open source 3D reconstruction of large models released by Tencent's WorldMirror team, which is an upgraded version of the WorldMirror series. It supports multi-view images, videos, and multi-modal a priori inputs such as camera position, internal reference, depth map, etc. It breaks through the traditional 3D reconstruction that only relies on...

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.

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...

MinerU2.5 is a decoupled visual language model jointly developed by Shanghai Artificial Intelligence Laboratory (AIL) and Peking University, focusing on efficiently processing high-resolution document image parsing. The core innovation lies in the two-phase design of "global layout detection followed by local content recognition": the first phase is a low-resolution...

LongCat-Audio-Codec is an open source speech codec solution from the LongCat team of Meituan. The program is designed for Speech Large Language Model (Speech LLM), through the semantic and acoustic dual Token parallel extraction mechanism , taking into account the semantic and acoustic features of speech ...

PaddleOCR-VL is Baidu's open source ultra-lightweight visual-language model, optimized for document parsing scenarios. The model contains only 0.9B parameters , through the fusion of dynamic high-resolution visual coder and lightweight ERNIE language model , while maintaining high accuracy and significantly reduce the computational overhead .

UniPixel is a novel multimodal model jointly proposed by Hong Kong Polytechnic University, Tencent, Chinese Academy of Sciences and Vivo to achieve pixel-level visual language understanding. By unifying object referencing and segmentation capabilities, it supports a variety of fine-grained tasks such as image segmentation, video segmentation, region understanding, and pi...

DiaMoE-TTS is a multi-dialect speech synthesis framework jointly open-sourced by Tsinghua University and Giant Network, based on the International Phonetic Alphabet (IPA), to solve the problems of dialect data scarcity, orthographic inconsistency, and complex phonological changes. Through a unified IPA front-end standardized phoneme representation to eliminate cross-dialect differences ...

Kandinsky 5.0 is the latest video generation model series developed by Russian AI team, focusing on lightweight design and high performance performance. The first model in the series, Kandinsky 5.0 Video Lite, has only 2 billion parameters but surpasses similar 14B models, especially...

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...

Pyscn is an intelligent code quality analysis tool designed for Python developers to detect potential problems in code to improve maintainability. It analyzes dead code through control flow diagrams, identifies duplicate code using APTED+LSH algorithm, calculates metrics such as module coupling and circle complexity...

Youtu-Embedding is a generalized text representation model open-sourced by Tencent's Youtu Lab, designed for enterprise-level applications. Through deep neural networks to map the text to a high-dimensional vector space, so that semantically similar sentences are closer in that space, to achieve accurate semantic retrieval.

SAIL-VL2 is an open source multimodal visual language model by the Byte Jump team, focusing on joint modeling of multimodal inputs such as images and text. Using the sparse mixture of experts (MoE) architecture and progressive training strategy, it achieves high performance at parameter scales from 2B to 8B, especially in the areas of graphic comprehension, math...

MineContext is an active context-aware AI partner open-sourced by the ByteDance Viking team to help users efficiently manage massive amounts of information and improve the efficiency of knowledge work. Over the screenshot and content understanding technology, automatically record the user's daily operations (such as browsing the web, editing documents, etc.), support...

nanochat is an open source project released by AI legend and former Tesla AI Director Andrej Karpathy that allows individuals to quickly train a small ChatGPT-like language model at a very low cost and simplicity. The entire project uses only about 800...

LLaVA-OneVision-1.5 is an open-source multimodal model by the EvolvingLMMS-Lab team, using 8B parameter scale, through a compact three-phase training process (language-image alignment, conceptual equalization and knowledge injection, and instruction fine-tuning) on 128 A800...

Paper2Video is an open-source presentation video project for automatic generation of academic papers by Show Lab at National University of Singapore. Using the PaperTalker multi-intelligence framework, papers are transformed into full presentation videos containing slides, subtitles, voiceover and speaker avatar...

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...

KAT-Dev-72B-Exp is an open-source programming-specific large language model launched by the Racer team, optimized based on reinforcement learning technology, which achieved an accuracy rate of 74.6% in the SWE-Bench Verified benchmark test, the best performance of any open-source model at present. The model uses innovative...

Jamba Reasoning 3B is a lightweight inference model open-sourced by Israeli AI startup AI21 Labs with strong performance and potential for a wide range of applications. It utilizes a hybrid SSM-Transformer architecture that combines Trans...

Agentic AI is the newest course on intelligent bodies launched by Ernest Ng.The course focuses on the design and construction of intelligent bodies, covering the four major design patterns of reflection, tool use, planning, and multi-intelligent body collaboration. Learners will master how to make intelligent bodies check outputs, autonomously adjust through theoretical explanations and code practice...

OpenAgents is the open source project that creates a network of AI agents and facilitates open collaboration between agents. A basic network infrastructure is provided to enable AI agents to seamlessly connect and collaborate. Users can quickly start their own agent network, extend functionality through a modular architecture, support...

Androidify is Google's open source project to help developers learn how to build AI-driven apps on Android. The project uses Google's latest technologies such as Jetpack Compose, Gemini API (via Fire...

Ling-1T is a trillion-parameter general-purpose language model open-sourced by Ant Group, which belongs to the flagship product of the Ling 2.0 series of Bering's large models. The model adopts a highly efficient MoE architecture, supports 128K context windows, and surpasses GPT in 7 benchmarks including code generation, mathematical reasoning, and logic test...

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...

Code2Video is innovative open source project that automatically converts code snippets into high quality video content (mp4 format). The project through a unique code-centric paradigm , the use of carbon-now-cli tools to generate code into beautiful images , the use of ffmpeg will be these ...

SceneGen is an open source method for generating 3D scenes from a single image at Shanghai Jiao Tong University. From a single scene image and a target resource mask, a complete scene containing multiple 3D resources is efficiently generated, including the geometric structure of the resources, texture and relative spatial location.

Ming-UniAudio is Ant Group's open source unified audio multimodal generation model that supports mixed input and output of text, audio, image and video. Using multi-scale Transformer and hybrid expert (MoE) architecture , through modality-aware routing mechanism to efficiently handle cross-modal ...

AIMangaStudio is a free AI manga creation tool that provides creators with a complete manga creation pipeline, including plot generation, sub-scene design, character setting and other functions, which can simplify the production process from script to manga page. It supports natural language generation of comic scripts, including plot, dialog...

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 .

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-1T-preview is an open source trillion-parameter big model of Ant Group, based on Ling 2.0 MoE architecture, pre-trained on 20T corpus, and trained in reasoning ability by self-developed reinforcement learning system ASystem. In natural language reasoning ...

RoboBrain-X0 is the world's first open source embodied model that supports zero-sample cross-ontology generalization open-sourced by Wisdom Source Research Institute, which is of great industrial significance. It can drive multiple real robots of different configurations to complete basic operation tasks without fine-tuning, and after a small amount of sample fine-tuning, it demonstrates the ability to replicate ...

Lynx is a high-fidelity personalized video generation model open-sourced by ByteDance that can generate identity-consistent videos with only a single portrait photo. Built on the diffusion Transformer (DiT) base model , the introduction of ID-adapter and Ref-adapte...

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...

DeepSeek-V3.2-Exp is a DeepSeek open source experimental AI model that significantly improves the efficiency of long text processing by introducing the DeepSeek Sparse Attention (DSA) mechanism. The model is based on DeepSeek...

HunyuanImage 3.0 (HunyuanImage 3.0) is a native multimodal image generation model released and open-sourced by Tencent. The model parameter size of 80B, is currently the best evaluation results, the largest number of parameters of the open source image generation model. Hybrid Image 3.0 supports real-time image generation, users can side...

Hunyuan3D-Part (Hybrid 3D-Part) is a 3D generation model released and open-sourced by Tencent. Composed of P3 - SAM and X - Part, it realizes high-precision and controllable component-based 3D generation for the first time, and supports 50 + components to be generated automatically. Users can first use...

AudioFly is KDDI open source AI model for text to generate sound effects. Based on the potential diffusion model architecture, with 1 billion parameters, trained on large-scale, diverse audio text datasets, covering AudioSet, AudioCaps, TUT and other public datasets and internal...

Hunyuan3D-Omni (Hybrid 3D-Omni) is an open source 3D asset generation framework by Tencent's Hybrid 3D team, which realizes accurate 3D model generation through multiple control signals. Based on Hunyuan3D 2.1 architecture, it introduces a unified control encoder that can handle point...

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...

CWM (Code World Model) is a 32-billion-parameter open-source world language model released by the Meta FAIR team, designed for code generation and reasoning. Introducing the concept of "world model", it can simulate the code execution process, predict the variable state changes, and advance...

Neovate Code is an open source intelligent programming assistant from Ant Group's Alipay Experience Technology Department, which improves development efficiency through artificial intelligence technology. With conversational development features, developers can describe the requirements through natural language, Neovate Code can understand and generate the corresponding generation...