🧰AI tool library

Claude Opus 4.7 is Anthropic's flagship model released in April 2026. It has 1M context, high-resolution visual understanding, self-verification and xhigh reasoning effort control, and is at the forefront of the industry in coding, agents and complex multi-step tasks.

Claude Opus 4.6 will be released in February 2026. Adaptive Thinking will be introduced to replace manual effort parameters and can independently determine the depth of reasoning.

Claude Sonnet 4.6, released in February 2026, is close to Opus-level coding and agent capabilities, but has more advantages in speed and cost.

Claude Sonnet 4.5 was released in September 2025 and is recommended by Anthropic as the default model.

This pioneering model easily transforms text and images into stunning, detailed, vividly colored visuals, especially for creators and marketers who require high-quality content

Flux AI Vincent graph model, the effect is comparable to Midjourney, crushing StableDiffusion

The most powerful open source MoE model DeepSeek-V3.2

DeepSeek latest model.

deepseek-r1 is a deep thinking model launched by deepseek.

DeepSeek-R1 based on Llama-3.3-70B, which uses reinforcement learning technology on a large scale during the training phase

DeepSeek-R1 based on Qwen2.5-32B, which uses reinforcement learning technology on a large scale during the training phase

DeepSeek-R1 based on Qwen2.5-Math-7B, which uses reinforcement learning technology on a large scale during the training phase

Deepseek-r1 is deployed in the open source version and can be searched online.

deepseek-r1

deepseek's latest large model, improved reasoning task performance, enhanced front-end development capabilities, upgraded Chinese writing, and optimized Chinese search capabilities

deepseek's latest large model, improved reasoning task performance, enhanced front-end development capabilities, upgraded Chinese writing, and optimized Chinese search capabilities

The default points to the latest deepseek-v3-1-terminus

Deepseek's newly launched hybrid reasoning model supports two reasoning modes: thinking and non-thinking, and is more efficient than deepseek-r1-0528.

DeepSeek-v3.1-fast is an extremely exciting breakthrough product designed to solve the problem of "speed and depth" of large models.

DeepSeek-V3.2 is the first model we launched that integrates thinking into tool use, and supports both thinking mode and non-thinking mode tool invocation.

DeepSeek-V3.2-Fast is a high-performance, low-latency variant based on the DeepSeek V3.2 architecture

DeepSeek-V4-Flash is a lightweight version of the DeepSeek V4 series. It focuses on high cost performance and high throughput efficiency. It is suitable for general conversations and basic text tasks. It also supports millions of Token long contexts and efficient reasoning.

DeepSeek-V4-Pro is a high-performance open source large model launched by DeepSeek. It has top reasoning and agent capabilities, supports ultra-long context, adapts to domestic Ascend chips, and is extremely cost-effective.

Google 2.0 launches latest model that outperforms state-of-the-art (O1) models in multi-modal AI capabilities

Gemini 2.5 Flash-Lite-Preview-09-2025 is a sub-model of the Gemini model family that focuses on ultra-low latency, high concurrency and the highest cost performance.

Gemini 2.5 Flash 09-2025 is the latest public preview version of the Gemini 2.5 Flash model launched by Google. It focuses on significantly improving the intelligence and application capabilities of the model while maintaining high efficiency and low cost.

Gemini 3 is Google’s smartest model series yet, powered by advanced inference capabilities.

gemini-3-flash-preview The smartest model we've ever built, combining speed and cutting-edge intelligence with exceptional search and grounding capabilities.

Nano Banana 2 delivers high-quality image generation and conversational editing with low latency at a mainstream price.

The most cost-effective model from Google, optimized for high-volume agent tasks, translation, and simple data processing.

Gemini 3.5 Flash has been officially released (GA), has stable performance and can be used in large-scale production environments.

The Gemini API provides text embedding models for generating embeddings for words, phrases, sentences, and code.

Gemini Flash Latest is an artificial intelligence model provided by google-vertex.

The latest model gemini-embedding-2-preview is the first multi-modal embedding model in the Gemini API.

gemma is a series of lightweight and advanced open source models launched by Google, built based on the research and technology of Google Gemini model.

gemma is a series of lightweight and advanced open source models launched by Google, built based on the research and technology of Google Gemini model.

Gemini Omni Flash is a multi-modal AI video model launched by Google.

Google's latest advanced artificial intelligence model, veo3.1 supports automatic audio generation for videos, with high quality and low price, the most cost-effective option. It supports first and last frames and Vincent videos (8 seconds in duration)

Google's latest advanced artificial intelligence model, veo3.1 high-quality mode, supports automatic audio generation for videos. The quality is extremely high and the price is extremely high. Please be careful when using it. It supports first and last frames and Vincent videos (8 seconds in duration).

Google's latest advanced artificial intelligence model, veo3.1 4k mode, supports automatic audio generation for videos, with high quality and low price, the most cost-effective option (8 seconds duration)

Google's latest advanced artificial intelligence model veo3.1-fast-components supports three cushion image inputs, which is faster, more stable and more controllable, redefining the creative freedom of Tusheng videos (8 seconds in duration)

Google's latest advanced artificial intelligence model, veo3.1 4k high-quality mode, supports automatic audio generation for videos. The quality is extremely high and the price is extremely high. Please be careful when using it. It supports first and last frames and Vincent videos (8 seconds in duration).

Google's latest advanced artificial intelligence model, veo_3_1 4k mode, supports automatic audio generation for videos, supports first frame delivery, does not support last frames, high quality and low price, the most cost-effective option (8 seconds duration)

Google's latest advanced artificial intelligence model, veo_3_1 4k mode, supports automatic audio generation for videos, supports first frame delivery, does not support last frames, high quality and low price, the most cost-effective option (8 seconds duration)

Google's latest advanced artificial intelligence model, veo3.1 fast +4k mode, supports automatic video and audio generation, high quality and low price, the most cost-effective option, supports first and last frames and Vincent videos (8 seconds in duration)

Google's latest advanced artificial intelligence model veo3.1-fast-components supports three cushion image inputs, which is faster, more stable, and more controllable. It redefines the creative freedom of Tusheng videos, 4k model (8 seconds in duration)

High-capacity Llama models suitable for complex analysis and forecasting tasks

High-capacity Llama models suitable for complex analysis and forecasting tasks

Llama-2-7B is the least parameterized but still powerful version of the entire Llama 2 family

The latest Meta Llama 3 model with 7 billion parameters

The latest Meta Llama 3 model with 700 million parameters

Llama 3rd generation 32K context model

Llama 3rd generation 32K context model

Meta’s 405B parameter Llama 3.1 model has powerful general capabilities

Meta's 405B parameter Llama 3.1 command fine-tuned version, more suitable for dialogue scenes

Meta’s 70B parameter Llama 3.1 model balances performance and efficiency

Meta's 70B parameter Llama 3.1 fast instruction version, faster response

Meta's 8B parameter Llama 3.1 model, a lightweight version suitable for resource-constrained scenarios

Meta's 11B parameter Llama 3.2 instruction version, lightweight dialogue model

Meta's 1B parameter Llama 3.2 instruction version, ultra-lightweight dialogue model

@cf/meta/llama-3.2-3b-instruct is an artificial intelligence model provided by cloudflare-workers-ai.

Meta's 90B parameter Llama 3.2 command fine-tuned version, more suitable for dialogue scenes

Meta-Llama-3-70B-Instruct is a fine-tuned version of the 70B parameter instruction. It is suitable for dialogue scenarios and performs better in understanding language details, context and performing complex tasks.

llama's best-in-class native multi-modal model delivers superior text for seamless long document analysis.

llama's best-in-class native multi-modal model delivers superior text, single H100 GPU efficiency, and 10M context windows for seamless long document analysis.

Midjourney blend mode

Midjourney custom scaling

Midjourney description mode

Midjourney high variation mode

Midjourneyimagination mode

Midjourney partial redraw

Midjourney low variation mode

Midjourney modal pattern

Midjourney pan mode

Midjourney regenerated

Midjourney upload service

Midjourney zoom mode

Midjourney variant mode

Midjourney zoom mode

Conch AI 02 brings your images to life through advanced AI motion synthesis technology.

MiniMax video model Hailuo 2.3 further upgrades the dynamic expressiveness based on the Hailuo 02 model, making the picture more realistic and stable.

MiniMax M2 lets the model generate dialogue content and tool calls based on the input context.

MiniMax-M2.5 has reached or refreshed the industry's SOTA in productivity scenarios such as programming, tool calling and searching, and office work.

MiniMax-M2.7 has reached or refreshed the industry's SOTA in productivity scenarios such as programming, tool calling and searching, and office work.

Conch sound design

Xiaomi MiMo-V2-Pro is designed for demanding real-world agency workflows.

MiMo-V2.5, a native full-modal model, supports text, image, video and audio understanding, and has powerful Agent capabilities

MiMo-V2.5-Pro ​​is oriented to complex task scenarios and is deeply adapted to Agent and Coding applications. It ranks first in the world's open source models on the GDPVal-AA and ClawEval lists.

Powerful multi-language programming capabilities, comprehensively upgrade the programming experience

Mistral Large is Mistral AI’s flagship model with top-level reasoning capabilities.

Mistral Large is Mistral AI’s flagship model with top-level reasoning capabilities.

Mistral Small is an optimized model that focuses on low latency and cost-effectiveness.

Replacement for GPT-3 ada and babbage base models.

Dynamic models in ChatGPT are continuously updated to the current version of GPT-4o.

codex-mini-2025-05-16 is a lightweight model-specific snapshot optimized for code generation and completion. It has both low latency and high cost performance, and is suitable for real-time programming assistance and high-concurrency code processing scenarios.

Generate a new version of the image DALL-E.

Replacement for GPT-3 curie and davinci base models.

Pure official high-speed GPT3.5 series, supports tools_call

Pure official high-speed GPT3.5 series, supports function_call

June 13, 2023 version of OpenAI GPT-3.5 Turbo 16K

Pure official high-speed GPT3.5 16K series, supports function_call

GPT-5 is our flagship model for cross-domain coding, reasoning, and agent tasks.

Purely official GPT4 series, 0613 series models support function_call

The latest gpt-4-0125-preview, an upgraded version of gpt-4-1106-preview, has stronger code generation capabilities, reduces model "laziness", and fixes non-English UTF-8 generation problems.

The latest gpt-4-1106-preview, also known as gpt-4-turbo, is 67% cheaper than gpt-4, supports 128k context, supports tools, and the knowledge deadline is April 2023

32K context window version of OpenAI GPT-4, updated June 13, 2023

Pure official GPT4 32K series

GPT All model, integrating official GPT-4, networking, image reading, drawing functions, and code interpreter

gpt-4-turbo-preview upgraded version, with stronger code generation capabilities, reducing model "laziness" and fixing non-English UTF-8 generation problems

All GPTs on the official website can be called

GPT-4.1 is an artificial intelligence model provided by openai.

GPT-4.1 mini is an artificial intelligence model provided by openai.

GPT-4.1 nano is an artificial intelligence model provided by openai.

openai latest model, gpt-4.5

GPT-4o is an artificial intelligence model provided by openai.

GPT All model, integrating official GPT-4o, networking, image reading, drawing functions, and code interpreter

GPT-4o mini is an artificial intelligence model provided by openai.

The Audio API allows developers to build speech-to-text and text-to-speech functionality.

gpt-5-all is the latest flagship model of openai

GPT-5 Chat points to the GPT-5 snapshot version currently used by ChatGPT.

GPT-5-Codex is a version of GPT-5 optimized for proxy encoding in Codex.

GPT-5 mini is a faster and more cost-effective version of GPT-5, especially suitable for processing clearly defined tasks and precise instructions.

GPT-5 Nano is our fastest and most affordable version of GPT-5, especially suited for summarization and classification tasks.

gpt-5 dedicated search model

GPT-5.1: Smarter, more conversational chat GPT, the most commonly used model, is now warmer, smarter, and better able to follow your commands.

GPT-5.1: Smarter, more conversational chat GPT, the most commonly used model, is now warmer, smarter and better at following your commands.

GPT-5.1-codex is a well-equipped heavy-duty main force.

GPT-5.1-codex-max breaks through the limits of context understanding and logical reasoning, and is designed to coordinate system-level reconstruction across warehouses, independently design and implement the full-stack ecosystem.

GPT-5.1-codex-mini An efficient model for rapid code generation and lightweight development tasks

GPT-5.2 is the best model for coding and intelligence tasks in a wide range of industries

The best model for coding and intelligence tasks in every industry

It is a god-level programming AI that can understand extremely complex system architectures, program entire software projects with zero errors, and even self-correct errors.

GPT-5.3-chat-latest is a high-speed dialogue model optimized by OpenAI. The response is more direct and natural, reducing preaching and refusal, and is suitable for daily lightweight tasks.

GPT-5.3-Codex redefines the role of AI in programming and pan-productivity through performance improvements, functional generalization, and security upgrades.

GPT-5.3-Codex-Spark Research Preview.

GPT-5.4 is our cutting-edge model for complex professional work.

GPT-5.4pro uses more computing resources to think deeper and provide always better answers.

GPT-5.4mini combines the advantages of GPT-5.4 into a faster, more efficient model designed for high-load workloads.

GPT‑5.4 nano is the lightest and fastest version of GPT‑5.4, designed for tasks where speed and cost are critical.

GPT-5.5 is the flagship large language model released by OpenAI on April 24, 2026. It is positioned as a new type of intelligence for practical work and agents. The core breakthrough lies in autonomous planning and execution of multi-step complex tasks. It is good at programming, computer operations, scientific research analysis and other fields, and does it with less T

GPT-5.5pro is now available for processing Responses API requests, including operating through the BBatch API to support multiple rounds of model interaction capabilities before responding to API requests, and will implement other advanced API features in the future.

Supports the reverse model of charging for all parameters of /v1/images/generations and /v1/images/edits

gpt-image-1-mini is our lightweight image generation model

GPT Image 1.5 is our latest image generation model, with better instruction tracking and prompt following.

Image generation model with better instruction tracking and prompt following.

gpt-image-2-all is our most advanced image generation model, supporting fast image generation and editing.

GPT OSS 120B is an artificial intelligence model provided by vultr.

The Gpt-oss-20b model achieves similar results to the OpenAI o3‑mini model on common benchmarks.

GPT-Reatime-1.5 is our flagship audio model for voice agents and customer support.

The full version of o1 will take more time to think (form a chain of ideas) to arrive at the answer, making it more suitable for performing complex reasoning tasks, especially scientific and mathematical tasks.

Spends more time thinking (forming chains of ideas) to arrive at an answer, making it better suited for complex reasoning tasks, especially science and math tasks

o1-mini is a fast, cost-effective inference model tailored for coding, math and science use cases.

<div>The o1-mini-all model integrates the basic capabilities of o1-mini, networking, image reading, drawing functions, and code interpreter. The file link can be placed anywhere in the prompt.</div>

o3 is an artificial intelligence model provided by openai.

OpenAI's most advanced reasoning model is good at multi-modal analysis and supports thinking functions

o3-deep-research is an artificial intelligence model provided by openai.

o3-mini is OpenAI’s latest small-scale inference model, optimized for coding, math, and scientific tasks.

o3-mini is openai’s latest thinking model

o3-mini-high is the latest thinking model of openai, which is slightly smarter than o3-mini

o4-mini is OpenAI’s latest efficient inference model, with low latency and high quality, and supports multi-modal input

o4-mini is OpenAI’s latest efficient inference model, with low latency and high quality, and supports multi-modal input

o4-mini-deep-research is our faster and more affordable deep research model, ideal for handling complex multi-step research tasks.

Now available in Sora 2, OpenAI's latest video generation model is more physically accurate, realistic, and easier to control than previous systems.

Reverse of sora-2, supports 10s, 15s, both are 720p

Sora 2 pro version, supports 4, 8, 12s, 720p and 1080p

The MiniMax speech large model can intelligently predict the emotion, intonation and other information of the text based on the context, and generate supernatural, high-fidelity, personalized speech.

The MiniMax speech large model can intelligently predict the emotion, intonation and other information of the text based on the context, and generate supernatural, high-fidelity, personalized speech.

The MiniMax speech large model can intelligently predict the emotion, intonation and other information of the text based on the context, and generate supernatural, high-fidelity, personalized speech.

The MiniMax speech large model can intelligently predict the emotion, intonation and other information of the text based on the context, and generate supernatural, high-fidelity, personalized speech.

The new generation voice HD model accurately restores the details of real tone and comprehensively improves the similarity of timbre.

New generation voice Turbo model, extremely fast response, vivid and natural tone expression

OpenAI’s most basic text generation model, the fastest but the weakest

OpenAI’s Babbage model for basic text processing tasks

A medium-capacity model of the OpenAI GPT-3 series, which is relatively balanced in speed and capability.

Based on GPT-3 text editing model, text can be modified according to instructions

OpenAI third-generation embedding model, more powerful than the small version, suitable for tasks requiring the highest performance

OpenAI third-generation embedding model, faster and cheaper than second-generation models, suitable for most tasks

Ada-based text embedding model optimized for various NLP tasks.

text-embedding-v1 is an efficient text vectorization basic model suitable for building semantic search, text clustering and recommendation systems.

The latest text moderation model for content security checks

Stable version text review model, providing reliable content review

<div>Text-to-speech model TTS supports setting timbre. The standard tts-1 model provides the lowest latency, but the quality is lower than the tts-1-hd model. </div>

<div>Text-to-speech model TTS supports setting timbre. The standard tts-1 model provides the lowest latency, but the quality is lower than the tts-1-hd model. </div>

Text-to-speech model TTS, supports setting timbre

Text-to-speech model TTS, supports setting timbre

Whisper can transcribe speech to text and translate multiple languages ​​into English

Different from video-based templates, this is a set of AI template gameplay with pictures as the core - just upload your photos to quickly generate an immersive AI content experience.

The Lipsync interface is specially designed to solve lip-syncing problems in videos.

Pixverse Mask Selection (or Swap Mask Generation) is an AI intelligent mask generation tool used by the PixVerse platform in the Modify function. It is specially used to accurately select the required mask in the video.

The Mimic function supports migrating the actions in the reference video to the target person picture to achieve action imitation and reconstruction.

The video editing (Modify) function supports editing any part of the existing video, including adding, replacing, deleting, modifying elements, or changing the overall style.

The multi-frame (Multi-transition) function allows you to generate a 1-30 second video by providing 2-7 key frames, ensuring a consistent style and smooth lens connection.

The Restyle feature instantly transforms your video clips into a new visual style—such as 3D, anime, cinematic, or augmented reality.

The sound effect generation (background music) interface can intelligently generate sound effects and environmental sounds that adapt to the video screen, and can also be input through text commands.

SWAP is a powerful video editing capability.

It belongs to the cutting-edge AI video foundation model. Its core capabilities include Text-to-Video (text to generate video) and Image-to-Video (picture to generate video). It supports uploading pictures or

The Grok 4 model is capable of deep reasoning and has been trained on xAI’s Colossus supercomputer, promising greater logical reasoning and text generation capabilities.

Grok 3 is an artificial intelligence model provided by xai.

grok-3-deepsearch deep network search model

Use grok3 to generate images

grok-3-reasoning is a model for grok to enhance reasoning capabilities

Grok 3 mini, which represents a new frontier in cost-effective reasoning capabilities.

grok-3-reasoning is a model for grok to enhance reasoning capabilities

grok-4-1-fast-non-reasoning is an AI model developed by xAI that is optimized for maximum speed when generating responses and executing agent tasks.

Grok 4.1 Fast is a top tool calling model under xAI, with 2 million context windows.

Grok 4.1 excels in creative, emotional and collaborative interactions, is able to capture subtle intentions more keenly, has a more engaging conversation experience, and maintains a high degree of consistency in personality traits, while fully inheriting the sharp intelligent performance and reliable performance of its predecessor.

Grok 4.20 is xai's latest flagship model, with industry-leading speed and proxy tool calling capabilities.

Grok 4.20 is xai's latest flagship model, with industry-leading speed and proxy tool calling capabilities.

grok-4-fast, the latest advancement in cost-effective inference models.

Grok 4 Fast (Non-Reasoning) is an artificial intelligence model provided by xai.

Grok-4.2 is an AI productivity system with trillions of parameters, 16-Agent cluster collaboration, supports real-time data stream processing, and can self-evolve through high-frequency weekly updates. It is specially optimized for complex decision-making, business automation, and personalized cognitive enhancement.

Grok-4.2-Fast is a fast and economical version of the Grok 4 series launched by xAI

Our most advanced flagship model, leading the industry in non-hallucination rates, agent tool calling and command following capabilities.

Grok-Imagine-Image is a multi-modal AI model launched by the X platform, which can generate high-quality images based on text descriptions.

Grok-Imagine-Image-Pro is an upgraded version of the multi-modal AI model of the X platform. It achieves higher-precision image generation and multiple rounds of optimization through stronger understanding and generation details.

grok's latest video model, 10s, supports audio and video simultaneous playback

MAI-DS-R1 is a DeepSeek-R1 inference model that was post-trained by the Microsoft AI team to fill information gaps in previous versions of the model and improve its risk signature while maintaining R1 inference capabilities.

doubao-seed-1.8 has stronger multi-modal understanding capabilities and Agent capabilities. The model can perform better in many complex tasks in the real world, further helping enterprises create value.

Supports adjusting the length of thinking, that is, supporting the reasoning_effort field in the API, which is divided into four modes: minimal, low, medium, and high.

The extremely fast response version is the extremely fast version in the Doubao Big Model 1.6 series. It has deep thinking and multi-modal understanding capabilities and supports 256K context.

Doubao-Seed-1.6-vision visual deep thinking model shows stronger general multi-modal understanding and reasoning capabilities in scenarios such as education, image review, inspection and security, and AI search question and answer.

Doubao-Seed-2.0-Code is optimized for enterprise-level programming needs. Based on the excellent Agent and VLM capabilities of Seed 2.0, it has especially enhanced coding capabilities. It not only has outstanding front-end capabilities, but also meets the common multi-language coding needs of enterprises.

The first full-modal understanding model in the Doubao model family supports native unified understanding of video, images, audio, and text, while upgrading Agent, Coding, and GUI capabilities.

Doubao-Seed-2.0-lite is a balanced model that balances performance and cost for high-frequency enterprise scenarios. Its comprehensive capabilities surpass the previous generation Doubao-Seed-1.8.

Doubao large model family full-modal understanding model, shorter thinking length, higher tokens efficiency

Doubao-Seed-2.0-mini provides the ultimate model inference speed for low-latency, high-concurrency and cost-sensitive scenarios.

Seedance 1.0 lite has powerful semantic understanding and command following capabilities, and can finely control the character's appearance, clothing style, and facial expressions. It also shows absolute advantages in multi-subject action analysis, embedded text response, degree adverbs, and lens switching response.

Seedance 1.0 lite has powerful semantic understanding and command following capabilities, and can finely control the character's appearance, clothing style, and facial expressions. It also shows absolute advantages in multi-subject action analysis, embedded text response, degree adverbs, and lens switching response.

Seedance 1.0 pro fast is a comprehensive model with bottom-end price and top-end performance, achieving an excellent balance between video generation quality, speed, and price.

Seedance 1.0 is the latest basic video generation model launched by the ByteDance Doubao model team.

Supports audio and video generated by Vincent, audio and video generated by first frame pictures, audio and video generated by first and last frames of pictures

Seedream 3.0 is a basic model for Chinese and English bilingual image generation that supports native high resolution. Its comprehensive capabilities are comparable to GPT-4o, ranking first in the world.

Doubao-Seedream-5.0-lite is the latest image creation model released by ByteDance.

Seedream 4.5 natively supports text, single image and multi-image input, enabling diverse gameplay methods such as multi-image fusion creation, image editing, and group image generation based on subject consistency, making image creation more free and controllable.

Kling's official subject-related services have been upgraded to a new version (supports 3s-8s)

Advanced lip-syncing model, including face recognition and lip-syncing functions.

AI audio generation model supports Wensheng sound effects, video dubbing and other functions.

The AI ​​image generation model supports functions such as text-generated images, image-generated images, multi-image reference images, and image editing.

Digital human video generation model that converts static human pictures into dynamic talking videos.

kling's creation subject (supports 3s-8s)

The custom timbre creation model allows users to create their own personalized voice timbre by uploading audio files or using audio from existing videos.

The AI ​​special effects center provides 136+ creative video special effects, including holiday themes, dynamic effects, style conversion, etc.

The intelligent image recognition model can perform content analysis and recognition on uploaded images, and extract key information, objects, scenes and other elements in the image.

Kling 2.6 Motion Control AI motion transfer technology - Real motion, precise control Transfer precise motions in reference videos to static character images.

The multi-element synthesis model supports the intelligent fusion of multiple elements in videos to achieve complex creative video production.

Kling-omni-image A complete creative suite designed for video and image creation KlingOmni combines video and image generation capabilities in a complete creative suite designed for professionals who need consistent high-quality output every time.

Keling O1, the product adheres to the Muti-modal language (MVL) concept, uses natural language as the semantic skeleton, and cooperates with multi-modal descriptions such as videos, pictures, and subjects to accurately understand your intentions and operate more intuitively.

The video extension model can intelligently extend existing videos to maintain picture coherence and style consistency, and is suitable for scenarios where video duration needs to be extended.

BAAI/bge-reranker-v2-m3 is a lightweight multi-language reranking model.

Zhipu AI general large model is suitable for scenarios that require high knowledge, reasoning ability, and creativity, such as advertising copywriting, novel writing, knowledge writing, code generation, etc.

GLM-5 is the new generation flagship base model of Zhipu. It is built for Agentic Engineering and can provide reliable productivity in complex system engineering and long-range Agent missions.

Zhipu AI's general large model provides more powerful question answering and text generation capabilities.

High cost performance: the most balanced model between reasoning power and price, context 128K, maximum output 4K

Extremely fast inference: ultra-fast inference speed and powerful inference effect, 8K context, maximum output 4K

Free call: Zhipu AI’s first free API, calling large models at zero cost.

Ultra-long input: specially designed for processing ultra-long text and memory-based tasks, context 1M, maximum output 4K

GLM-4.5 and GLM-4.5-Air are the latest flagship model series, basic models designed for intelligent agent applications.

GLM-4.5 and GLM-4.5-Air are the latest flagship model series, basic models designed for intelligent agent applications.

GLM-4.5 and GLM-4.5-Air are the latest flagship model series, basic models designed for intelligent agent applications.

GLM-4.5 and GLM-4.5-Air are the latest flagship model series, basic models designed for intelligent agent applications.

GLM-4.5 and GLM-4.5-Air are the latest flagship model series, basic models designed for intelligent agent applications.

GLM-4.6 is an artificial intelligence model provided by zhipuai-coding-plan.

The GLM-4.6V series is an important iteration of the GLM series in the multi-modal direction. It increases the context window during training to 128k tokens, achieves the same parameter scale SOTA in visual understanding accuracy, and for the first time integrates Function in the model architecture.

GLM-4.7 is the latest flagship model of GLM. GLM-4.7 strengthens coding capabilities, long-term task planning and tool collaboration for agentic coding scenarios, and has achieved leading performance among open source models in the current lists of multiple public benchmarks.

GLM-5.1 is a model designed for Long Horizon Tasks.

Kimi-K2 is a MoE architecture basic model launched by Moonshot AI with super strong code and agent capabilities. It has a total parameter of 1T and an activation parameter of 32B.

Kimi K2 is a breakthrough hybrid expert model designed for superior performance in cutting-edge knowledge, reasoning and programming tasks.

kimi-k2 is a MoE architecture basic model with strong code and agent capabilities.

kimi-k2.5 is Kimi's most versatile model to date, with a native multi-modal architecture design that supports both visual and text input, thinking and non-thinking modes, dialogue and agent tasks.

Kimi K2.6 is Kimi's latest and most intelligent model. It has stronger and more stable long-range code writing capabilities, significantly improved command following and self-correction capabilities, and supports text, image and video input, thinking and non-thinking modes, dialogue and Agent tasks.

Moonshot AI has launched a language model with hundreds of billions of parameters, which has excellent semantic understanding, instruction following and text generation capabilities.

Moonshot AI has launched a language model with hundreds of billions of parameters, which has excellent semantic understanding, instruction following and text generation capabilities.

Moonshot AI has launched a language model with hundreds of billions of parameters, which has excellent semantic understanding, instruction following and text generation capabilities.

This interface is used to input text to generate corresponding reading audio, and can control the speaking speed, volume, and emotion of the reading.

This interface is used to input text to generate corresponding reading audio, and can control the speaking speed, volume, and emotion of the reading.

vidu2.0 has a fast generation speed, supports 360p, 720p, 1080p, supports picture-based videos, reference videos, first and last frames

The video generated by viduq1 has a clear picture, supports 1080p, supports Vincent videos, Picture videos, Reference videos, first and last frames

viduq1-classic, rich and stable camera operation, supports graphic videos and first and last frames, and supports 1080p

viduq2 supports Vincent videos and reference videos, with good dynamic effects and rich generated details. It also supports Vincent pictures, picture editing, and reference pictures.

viduq2-pro supports reference video, video editing, 540p, 720p, 1080p

viduq3 has strong picture quality, supports intelligent lens cutting, good dynamic effects, and the strongest balance. It supports 540p, 720p, and 1080p.

viduq3-pro supports audio and video simultaneous playback, supports intelligent mirror cutting, and supports 720p, 1080p

viduq3-mix supports intelligent lens cutting, and the consistency of multiple cameras is better. Supports 720p, 1080p

ERNIE 3.5 is Baidu's self-developed flagship large-scale language model, covering a massive amount of Chinese and English corpora. It has powerful general capabilities and can meet the requirements of most dialogue Q&A, creative generation, and plug-in application scenarios; ensuring the timeliness of Q&A information.

The most powerful language model in Baidu Wenxin series, its understanding, generation, logic, and memory capabilities have reached the top level in the industry.

Baidu's self-developed vertical scene large language model is suitable for application scenarios such as game NPCs, customer service dialogues, and dialogue role-playing. The character style is more distinctive and consistent, the ability to follow instructions is stronger, and the reasoning performance is better.

ERNIE Functions 8K is a powerful AI model developed by Baidu that is designed for advanced natural language processing tasks and can be integrated with external functions to enhance its functionality.

Baidu's self-developed lightweight large language model takes into account excellent model effects and reasoning performance, and is suitable for low computing power AI accelerator card reasoning.

Baidu's self-developed lightweight large language model takes into account excellent model effects and reasoning performance, and is suitable for low computing power AI accelerator card reasoning.

Baidu's self-developed lightweight large language model takes into account excellent model effects and reasoning performance, and is suitable for low computing power AI accelerator card reasoning.

ERNIE-Tiny-8K is a lightweight Chinese pre-trained language model developed by the Baidu team.

Embedding-V1 is a text representation model based on Baidu Wenxin large model technology. It converts text into a vector form represented by numerical values ​​and can be used in text retrieval, information recommendation, knowledge mining and other scenarios.

It is a new generation cognitive intelligence large model launched by iFlytek. It has cross-domain knowledge and language understanding capabilities. It can understand and perform tasks based on natural dialogue, and provides language understanding, knowledge Q&A, logical reasoning, mathematical problem solving, and code understanding.

It is a new generation cognitive intelligence large model launched by iFlytek. It has cross-domain knowledge and language understanding capabilities. It can understand and perform tasks based on natural dialogue, and provides language understanding, knowledge Q&A, logical reasoning, mathematical problem solving, and code understanding.

It is a new generation cognitive intelligence large model launched by iFlytek. It has cross-domain knowledge and language understanding capabilities. It can understand and perform tasks based on natural dialogue, and provides language understanding, knowledge Q&A, logical reasoning, mathematical problem solving, and code understanding.

It is a new generation cognitive intelligence large model launched by iFlytek. It has cross-domain knowledge and language understanding capabilities. It can understand and perform tasks based on natural dialogue, and provides language understanding, knowledge Q&A, logical reasoning, mathematical problem solving, and code understanding.

bce-reranker-base_v1 is a bilingual and cross-language reranking model developed by NetEase Youdao, supporting Chinese, English, Japanese and Korean.

Qwen3-Reranker-0.6B is a text reranking model from the Qwen3 series.

HappyHorse-1.0-I2V supports Tusheng videos, has highly restored dynamic picture generation capabilities, can accurately understand text semantics, and outputs smooth, natural, high-quality videos with rich details.

HappyHorse-1.0-R2V supports reference raw videos, more stable subject and scene reference, and supports up to 9 picture references, which can accurately maintain creative intentions and achieve stronger performance capabilities.

HappyHorse-1.0-T2V supports Wensheng video, has the ability to generate highly restored dynamic pictures, can accurately understand text semantics, and output smooth, natural, and high-quality videos with rich details.

HappyHorse-1.0-Video-Edit supports video editing and natural language commands to edit videos. It can refer to up to 5 pictures to edit video elements locally or globally. It can accurately reproduce the dynamic process of the video and achieve stronger performance capabilities.

Tongyi Qianwen QVQ visual reasoning model supports visual input and thinking chain output, showing stronger capabilities in mathematics, programming, visual analysis, creation and general tasks.

qwen-image-edit-2509 is the latest painting model

Tongyi Qianwen's billion-level ultra-large-scale language model supports input in different languages ​​such as Chinese and English.

The Qwen-Image-2.0 series accelerated version model realizes the integration of image generation and image editing; it has more professional text rendering 1k token command support capabilities, more delicate real texture, delicate depiction of realistic scenes, and stronger semantic compliance capabilities.

Tongyi Qianwen long text model

Tongyi Qianwen's billion-level ultra-large-scale language model supports input in different languages ​​such as Chinese and English.

Based on the comprehensively upgraded flagship translation model of Qwen3, it supports translation into 92 languages. The model performance and translation effect are fully upgraded, providing more stable terminology customization, format reduction, and domain prompt capabilities, making the translation more accurate and natural.

Based on the fully upgraded lightweight text translation model of Qwen3, it supports 92 language translations. The model performance and translation effect are fully upgraded, providing more stable terminology customization, format reduction, and domain prompt capabilities, making the translation more accurate and natural.

The stable version of Tongyi Qianwen Omni's new multi-modal understanding and generation large model supports text, image, voice and video input, and outputs text, providing 4 natural dialogue sounds.

That is, the Tongyi Qianwen ultra-large-scale visual language model.

The Tongyi Qianwen VL-PLUS model is relatively balanced in terms of effect and cost. If you are not sure about using a certain model for the time being, you can try the Tongyi Qianwen VL-PLUS model first.

Tongyi Qianwen 1.5 generation 110B parameter dialogue model

Tongyi Qianwen 1.5 generation 14B parameter dialogue model

Tongyi Qianwen 1.5 is an open source chat model with 32B scale parameters aligned with human instructions.

Tongyi Qianwen 1.5 generation 7B parameter dialogue model

Tongyi Qianwen 2 is a 1.5B scale model that is open source to the outside world.

Tongyi Qianwen 2 is an open source MOE model with 57B scale and 14B activation parameters.

Tongyi Qianwen 2nd generation 7B parameter instruction model

Tongyi Qianwen 2-VL-72B, extends the context to 32k, enhances image understanding capabilities, and can better recognize multi-language and handwriting in pictures.

Tongyi Qianwen 2-VL-72B, extends the context to 32k, enhances image understanding capabilities, and can better recognize multi-language and handwriting in pictures.

Qwen2.5 series 14B model, compared with Qwen2, Qwen2.5 has gained significantly more knowledge, and has been greatly improved in programming ability and mathematical ability.

Qwen2.5 series 14B model, compared with Qwen2, Qwen2.5 has gained significantly more knowledge, and has been greatly improved in programming ability and mathematical ability.

Tongyi Qianwen 2.5 generation 32B parameter instruction model

Qwen2.5 series 3B model, compared with Qwen2, Qwen2.5 has gained significantly more knowledge, and has been greatly improved in programming ability and mathematical ability.

Qwen2.5 series 72B model, compared with Qwen2, Qwen2.5 has gained significantly more knowledge and has been greatly improved in programming ability and mathematical ability.

qwen2.5-7b-instruct is an artificial intelligence model provided by aliyun-bailian.

qwen2.5-7b-instruct-1m is an artificial intelligence model provided by aliyun-bailian.

Qwen2.5 series programming expert 14B model, compared with Qwen2, Qwen2.5 has gained significantly more knowledge and has greatly improved in programming ability and mathematical ability.

Qwen2.5 series programming expert 32B model, compared with Qwen2, Qwen2.5 has gained significantly more knowledge and has greatly improved in programming ability and mathematical ability.

Qwen2.5 series programming expert 7B model, compared with Qwen2, Qwen2.5 has gained significantly more knowledge and has greatly improved in programming ability and mathematical ability.

Qwen2.5 series mathematics expert 72B model, compared with Qwen2, Qwen2.5 has gained significantly more knowledge, and has been greatly improved in programming ability and mathematical ability.

Qwen2.5 series mathematics expert 7B model, compared with Qwen2, Qwen2.5 has gained significantly more knowledge, and has been greatly improved in programming ability and mathematical ability.

The flagship visual language model of the Qwen model family has achieved a huge leap forward compared to the previously released Qwen2-VL.

Command following, mathematics, problem solving, and overall coding have been improved, and the ability to recognize everything has been improved. It supports multiple formats to directly and accurately locate visual elements. It supports the understanding of long video files (up to 10 minutes) and the positioning of events at the second level. It can understand the sequence and speed of time, and supports control based on analysis and positioning capabilities.

qwen2.5-vl-72b-instruct is an artificial intelligence model provided by aliyun-bailian.

Command following, mathematics, problem solving, and overall coding have been improved, and the ability to recognize everything has been improved. It supports multiple formats to directly and accurately locate visual elements. It supports the understanding of long video files (up to 10 minutes) and the positioning of events at the second level. It can understand the sequence and speed of time, and supports control based on analysis and positioning capabilities.

The data set of Qwen3 is significantly expanded compared to Qwen2.5.

The data set of Qwen3 is significantly expanded compared to Qwen2.5.

Achieve effective integration of thinking mode and non-thinking mode, and switch modes during the conversation.

The qwen3-235b-a22b-instruct-2507 model released in July 2025 only supports non-thinking mode and is an upgraded version of qwen3-235b-a22b (non-thinking mode).

The data set of Qwen3 is significantly expanded compared to Qwen2.5.

The qwen3-30b-a3b-instruct-2507 model released in July 2025 only supports non-thinking mode and is an upgraded version of qwen3-30b-a3b (non-thinking mode).

The qwen3-30b-a3b-thinking-2507 model released in July 2025 only supports thinking mode and is an upgraded version of qwen3-30b-a3b (thinking mode).

The data set of Qwen3 is significantly expanded compared to Qwen2.5.

The data set of Qwen3 is significantly expanded compared to Qwen2.5.

The data set of Qwen3 is significantly expanded compared to Qwen2.5.

The data set of Qwen3 is significantly expanded compared to Qwen2.5.

Based on the code generation model of Qwen3, it has powerful coding agent capabilities, is good at tool calling and environment interaction, and can achieve independent programming, excellent coding capabilities and general capabilities.

Based on the code generation model of Qwen3, it inherits the coding agent capabilities of Qwen3-Coder-Plus, supports multiple rounds of tool interaction, and focuses on optimizing warehouse-level understanding capabilities and increasing tool calling stability.

qwen3-coder-30b-a3b-instruct is a code generation model based on Qwen3. It has powerful Coding Agent capabilities and is good at tool calling and environment interaction. It can achieve independent programming and excellent coding capabilities while being versatile.

Tongyi Qianwen code model open source version.

Tongyi Qianwen 3 series Max model, compared with the snapshot on September 23, 2025, this version realizes the effective integration of thinking mode and non-thinking mode, and the overall effect of the model has been greatly improved in all aspects.

Based on the text ranking model trained by Qwen LLM base, the input Query and candidate Docs are sorted by relevance. It supports 100+ languages ​​and long text input. It is suitable for text retrieval, RAG and other scenarios. The effect is aligned with the open source Qwen3-Rerank series model.

Reverse version of qwen3-max-preview

Qwen3-Next 80B-A3B Instruct is an artificial intelligence model provided by alibaba-cn.

qwen3-next-80b-a3b-thinking, released in September 2025, only supports thinking mode. Compared with qwen3-235b-a22b-thinking-2507, it has improved command following capabilities and the summary reply is more streamlined.

Alibaba Cloud's Tongyi Qianwen VL open source version.

Alibaba Cloud's Tongyi Qianwen VL open source version.

The Instruct version of the second largest MoE model in the Qwen3-VL series has fast response speed and supports ultra-long contexts such as long videos and long documents; it has comprehensively upgraded image/video understanding, spatial perception and all-things recognition capabilities; it has visual 2D/3D positioning capabilities and is capable of complex real-world tasks.

Qwen3 - Thinking version of the second largest MoE model in the VL series, with fast response speed, stronger multi-modal understanding and reasoning, visual agent, long video and long document and other ultra-long context support capabilities; comprehensive upgrade of image/video understanding, spatial perception and everything recognition capabilities, capable of complex tasks

The non-inference version of the largest Dense model of the Qwen3-VL series, its overall performance is second only to Qwen3-VL-235B-Instruct. It has excellent document recognition and understanding capabilities, strong spatial perception and recognition of all things, and visual 2D detection/spatial reasoning capabilities reaching SO

The reasoning version of the largest Dense model of the Qwen3-VL series. Its multi-modal reasoning ability is second only to Qwen3-VL-235B-Thinking. It has outstanding STEM & mathematics problem-solving ability, general image and video understanding ability, and its multi-modal Agent ability reaches S

The Instruct version of the Qwen3-VL series 8B Dense model takes up less video memory, comprehensively upgrades image/video understanding, long video and long document and other ultra-long context support, spatial perception and everything recognition capabilities, and is capable of complex real-world tasks.

The Thinking version of the Qwen3-VL series 8B Dense model takes up less video memory and can complete multi-modal understanding and reasoning; supports long videos and long documents and other long contexts, visual 2D/3D positioning; comprehensively upgrades image/video understanding, spatial perception and all things recognition capabilities

The Qwen3 series of small-size visual understanding models realize the effective integration of thinking mode and non-thinking mode. The effect is better than the open source version Qwen3-VL-30B-A3B, and the response speed is fast.

Tongyi Qianwen VL is a text generation model with visual (image) understanding capabilities. It can not only perform OCR (picture text recognition), but also further summarize and reason, such as extracting attributes from product photos, solving problems based on exercise diagrams, etc.

The Qwen3.5 series 122B-A10B native visual language model is based on a hybrid architecture design and integrates a linear attention mechanism and a sparse hybrid expert model to achieve higher reasoning efficiency.

Qwen3.5 series 27B native visual language Dense model incorporates a linear attention mechanism; it has fast response speed and combines inference speed and performance.

The Qwen3.5 series 35B-A3B native visual language model is based on a hybrid architecture design and integrates a linear attention mechanism and a sparse hybrid expert model to achieve higher reasoning efficiency.

The Qwen3.5 series 397B-A17B native visual language model is based on a hybrid architecture design and integrates a linear attention mechanism and a sparse hybrid expert model to achieve higher reasoning efficiency.

The wen3.5 native visual language series Plus model is based on a hybrid architecture design that integrates a linear attention mechanism and a sparse hybrid expert model to achieve higher reasoning efficiency.

Qwen3.6 series 27B native visual language Dense model. Compared with 3.5-27B, the model effect focuses on improving agentic coding capabilities, model STEM and reasoning capabilities. In terms of visual modality, it has spatial intelligence, object positioning and detection capabilities.

The Qwen3.6 series 35B-A3B native visual language model is based on a hybrid architecture design and integrates a linear attention mechanism and a sparse hybrid expert model to achieve higher reasoning efficiency.

The Max model Preview version, which is the largest and most comprehensive in the Qwen3.6 series, currently has the pure text model capability open for experience.

The Qwen 3.6 native visual language series Plus model shows excellent performance comparable to the current top cutting-edge models, and the model effect is significantly improved compared to the 3.5 series.

The Max model, which is the largest and most comprehensive model in the Qwen3.7 series, currently has pure text model capabilities open for experience.

QwQ - open source version.

The QwQ-Preview model is an experimental research model developed by the Qwen team in 2024, focusing on enhancing AI reasoning capabilities, especially in the fields of mathematics and programming.

The stable version of QwQ is based on the QwQ reasoning model trained by the Qwen2.5 model, which greatly improves the model's reasoning capabilities through reinforcement learning.

wan2.5-i2v-preview is an artificial intelligence model provided by aliyun-bailian.

Wanxiang 2.6.

Wanxiang 2.6-Tusheng Video-Flash, faster and more cost-effective generation.

Wanxiang 2.7 - the ultimate model for image generation and editing, supports Vincentian pictures, Vincentian group pictures, Tusheng group pictures, image editing, multi-image reference generation, interactive editing, and has stronger performance in text rendering, subject consistency, and complex instruction compliance.

Z-Image-Turbo is an efficient image generation model that ranked first in the world among Vincentian open source models in the Artificial Analysis evaluation. It can generate photorealistic images comparable to large-scale commercial models with only 6 billion parameters and 8 steps of reasoning. It is also used in both Chinese and English.