Home/AI directory

103 AI sites

Every site here is one people actually use — not a scraped pile of links. Each entry says what it does and who it suits, and flags the traps: free-tier limits and overseas access.

Common
12 categories · 103 sites · updated 2026-11-03

Chat assistants

12

Where you talk to a model. The same model behaves differently in different products — system prompts and tooling matter more than the parameter count.

ChatGPTHotOpenAI's general chat entry point. The GPT-6 line launches here, with Canvas, web access and code execution in one place.Broadest ecosystemFreeOverseas↗ClaudeHotAnthropic's chat product. The best reputation for long-document handling and writing quality, and Artifacts renders a previewable page on the spot.Best for long textFreeOverseas↗GeminiHotGoogle's multimodal assistant, wired into Search, Workspace and Android. Very long context is its home turf.MultimodalFreeOverseas↗DeepSeekHotDeepSeek's official chat. The open weights and the API are the same model, the reasoning trace can be expanded, and prices are among the lowest in its tier.ValueFree↗DoubaoHotByteDance's chat product. The most complete voice interaction and persona features, and the easiest for ordinary users to pick up.Strong voiceFree↗KimiMoonshot's long-context assistant. Whole-page parsing and batch file handling are smooth — built for grinding through reports and papers.Long contextFree↗QwenAlibaba Cloud's chat entry point, where you can try the whole Qwen size range. Attachment parsing and document translation are the common uses.Full stackFree↗ChatGLMZhipu's chat product. The GLM line reads naturally in Chinese, and agent and tool-call configuration is fairly open.Natural ChineseFree↗Tencent YuanbaoRuns both Hunyuan and DeepSeek with a switch between them; its WeChat Official Account search is a corpus you cannot get elsewhere.WeChat corpusFree↗Le ChatMistral's chat product. Response speed is the selling point, and European compliance often makes it the default choice.FastFreeOverseas↗GrokxAI's assistant. Dense real-time information and a looser tone — good for reading public sentiment and how a story is developing.Real-timeFreeOverseas↗PoeCompare dozens of models and community bots under one subscription — the least painful place to run model-selection experiments.Multi-modelFreeOverseas↗

Coding & dev

11

From single-line completion to whole-repo rewrites — the levels differ enormously. Decide whether you want a line completed or a module refactored.

CursorHotA VS Code-based AI editor with whole-repo indexing and multi-file rewrites; the gap to a completion plugin is cross-file refactoring.Repo-wide editsFreeOverseas↗GitHub CopilotHotThe most widely adopted completion tool, already tied into PRs, code review and issues — the least resistance when rolling out to a team.Team standardFreeOverseas↗WindsurfAn editor driven by the Cascade flow that chains multiple edits together — built for letting the AI run the whole path.Agent flowFreeOverseas↗ClineAn open-source VS Code agent with your own API key. Every step needs your approval — control in exchange for peace of mind.Open, controllableFreeOverseas↗ContinueAn IDE plugin that can point at local models — the usual answer when code cannot leave the corporate network.Local modelsFreeOverseas↗TraeByteDance's AI IDE with a Chinese interface and direct domestic access — a lighter alternative to Cursor.Direct in ChinaFree↗LingmaAlibaba Cloud's coding assistant; the enterprise edition can be self-hosted, and its Java completion is among the best in China.Private deployFree↗v0By Vercel: one sentence produces a working React/Tailwind page — the fastest route to a front-end prototype.Instant front-endFreeOverseas↗Bolt.newRuns a full project in the browser and deploys as soon as you finish editing — for validating ideas, not long-term maintenance.Full stack in browserFreeOverseas↗ReplitAn online IDE with an agent that goes from writing code to deploying on one page — handy for teaching and group projects.All-in-oneFreeOverseas↗DevinAn autonomous agent sold as an AI engineer: it reads docs and opens PRs, though a human still needs to watch the critical steps.AutonomousOverseas↗

Image generation

10

Text-to-image and image editing. Chinese tools generally understand Chinese prompts better than overseas ones.

MidjourneyHotStill in the top tier for aesthetics — style consistency and mood are its steadiest strengths, useful for concepts and visual reference.Top-tier aestheticsOverseas↗JimengHotThe image and video tool from the CapCut ecosystem: good Chinese prompt comprehension, fast output, and it plugs into an editing workflow.Chinese-friendlyFree↗KlingHotFrom Kuaishou: the strongest domestic image-to-video on motion plausibility and physical feel, and it does image generation too.Top domestic videoFree↗LiblibAIChina's largest model and workflow community, with a huge library of LoRA and ComfyUI workflows you can run online.Model hubFree↗Stability AIThe official Stable Diffusion home. Open weights you can deploy locally, with a level of control and ecosystem no closed product offers.Open, localFreeOverseas↗FLUXBlack Forest Labs' model: strong text rendering and photoreal detail, and it sits under many commercial products.Strong detailFreeOverseas↗IdeogramText inside the image almost never garbles — first choice for posters, covers and anything with a tagline.Reliable textFreeOverseas↗RecraftBuilt for vector work and brand visuals, exporting SVG so recolouring or retexting needs no redraw.SVG outputFreeOverseas↗Leonardo.AiA generation platform aimed at game art: model training, inpainting and upscaling in one chain, so assets come out fast.Game artFreeOverseas↗KreaA live canvas that renders as you draw — better for finding composition and lighting than rolling the dice on prompts.Live canvasFreeOverseas↗

Video generation

8

Text-to-video, image-to-video and digital humans. Consistency is still the bottleneck — finished clips mostly come down to how many times you roll.

Audio & music

7

Speech synthesis, voice cloning and music generation. The gap between Chinese and overseas tools on Chinese prosody and emotion is closing.

Writing & office

8

Documents, spreadsheets, meeting notes, email. The value here is fitting into an existing workflow, not how strong the model is.

Agents & automation

8

Models calling tools and running processes on their own. Whether it can finish ten steps reliably is what separates a toy from a tool.

Design & creation

6

Posters, decks, UI and brand assets. AI drafts fast; a human still has to close it out.

Research

8

Literature search, paper skimming and citation tracing. Run through these before writing a review and save hours of filtering.

Models & infrastructure

10

Model hosting, inference services and routing gateways. The line where self-hosting beats calling an API is lower than you think.

Hugging FaceHotHome to open models and datasets, with model cards, inference endpoints and Spaces demos in one place — the first stop when choosing a model.Open model hubFreeOverseas↗OpenRouterHotOne key for hundreds of models with live price and availability comparison — the easiest way to set up multi-model fallback.One key, all modelsFreeOverseas↗SiliconFlowHotA Chinese inference provider with cheap, stable open-model APIs; migration is just changing the base_url on an OpenAI-compatible client.Cheap in ChinaFree↗OllamaHotOpen models on your own machine in one command — the standard for development and privacy-sensitive work, and best on a Mac.One-command localFreeOverseas↗ModelScopeAlibaba's open model community with fast domestic downloads, and a free Notebook allowance that covers small models.Fast in ChinaFree↗ReplicatePay-per-call open models, with image and video models added fastest — for validation rather than long-term production.Pay per callFreeOverseas↗Together AIHosted inference and fine-tuning for open models, with throughput and pricing that pay off at volume.Hosted fine-tuningFreeOverseas↗GroqCustom LPU silicon makes inference unusually fast — the first choice for real-time voice and low-latency work.Very fastFreeOverseas↗vLLMA high-throughput inference engine; PagedAttention squeezes the most out of VRAM, which makes it the default for self-hosted serving.Self-hosted servingFreeOverseas↗Alibaba Cloud BailianThe API platform for the whole Qwen family, with enterprise quotas, auditing and private knowledge bases in one place.Enterprise APIFree↗

News & community

7

See how others use these models and where they get stuck. First-hand news on new releases usually lands here first.

Links point to the official site; prices and free allowances change without notice.“Overseas”marks a service hosted abroad, so connecting from mainland China may need a proxy — it does not mean it is unavailable. Found a dead link or want to suggest a site? Email jiaoliudao@qq.com.