DeepSeek, Qwen (Alibaba), GLM (Zhipu/Z.ai), and Kimi (Moonshot AI) are the four Chinese AI labs whose models you can use free of cost for coding in 2026 — through free web chat interfaces (chat.deepseek.com, chat.qwen.ai, chat.z.ai, kimi.ai), free-tier models on OpenRouter, or by self-hosting the open weights under permissive MIT/Apache 2.0 licenses. Each one has a different sweet spot: DeepSeek is the cheapest all-round default, Qwen3-Coder is the best pick for local/self-hosted work with huge context windows, GLM is tuned for agentic coding tools like Claude Code and Cline, and Kimi K2 is built for long, multi-step coding sessions.
Why Chinese AI labs are the free option right now
Every major Chinese AI lab — DeepSeek, Alibaba (Qwen), Zhipu/Z.ai (GLM), and Moonshot AI (Kimi) — ships both a free hosted chat product and open-weight model files under permissive licenses (MIT or Apache 2.0 in most cases). That combination doesn't really exist among the US labs: OpenAI, Anthropic, and Google keep their frontier coding models closed and metered. If your goal is "free, usable for real coding work, no credit card," the practical shortlist in 2026 is DeepSeek, Qwen, GLM, and Kimi.
DeepSeek — the default free choice for coding
DeepSeek's own chat product at chat.deepseek.com is free to use in the browser,
including its DeepThink reasoning mode, and handles code generation, debugging, and file reading
without a paywall. If you want programmatic access instead of the chat UI, DeepSeek V3 and DeepSeek
V3.1 are both listed on OpenRouter as zero-cost models (deepseek/deepseek-chat-v3-0324:free
and deepseek/deepseek-chat-v3.1:free), so you can wire them into an IDE extension or agent
harness without paying per token. If you outgrow the free tier, DeepSeek's own paid API for V3.2 is
still among the cheapest frontier-class options at roughly $0.21 per million input tokens and $0.31 per
million output tokens. The weights themselves are released under an MIT license, so self-hosting is
also fully free if you have the GPU budget.
Qwen (Alibaba) — Qwen3-Coder and Qwen Chat
Alibaba's Qwen Chat at chat.qwen.ai is free with no usage caps and includes a
dedicated Qwen-Coder mode alongside the general chat model, so you don't need an API key just to get
coding help. For developers who want an API, the open-weight Qwen3-Coder model (480B parameters,
35B active, mixture-of-experts) is listed for free on OpenRouter as qwen/qwen3-coder:free,
with a native context window in the hundreds of thousands of tokens — useful for feeding in large
codebases. One caveat: Alibaba's own free OAuth quota for the Qwen Code CLI (once 1,000 free requests/day)
was cut down and then closed entirely in April 2026 in favor of a paid $50/month Coding Plan. If you want
Qwen for CLI-based agent workflows specifically, budget for that plan or route through a free OpenRouter
model instead. The weights are Apache 2.0 licensed, so self-hosting Qwen3-Coder is also an option.
GLM (Zhipu / Z.ai) — built for agentic coding tools
Zhipu's consumer product, Z.ai, offers a free chat experience at chat.z.ai with no sign-up and no card required, running its GLM line of models directly in the browser. GLM-4.6, released under an MIT license, is a 357-billion-parameter model with a context window expanded from 128K to 200K tokens, and it's specifically noted for strong real-world performance inside agentic coding tools such as Claude Code, Cline, Roo Code, and Kilo Code — i.e., it's designed to be dropped into the same tool-calling workflows developers already use with proprietary models, not just chatted with directly. Because the weights are MIT-licensed, GLM is also one of the cleanest options to self-host if you'd rather not depend on a hosted free tier that could change later.
Kimi K2 (Moonshot AI) — best for long, multi-step coding runs
Moonshot AI's Kimi chat product at kimi.ai is free to use in the browser. The underlying K2 model family is aimed specifically at long-horizon, multi-step coding tasks across languages like Python, Rust, and Go, including turning prompts or visual mockups into working interfaces — the kind of extended, multi-turn agent session where cheaper models tend to lose the thread. Moonshot doesn't currently run a permanent free tier for its hosted production API, so the genuinely free path for programmatic access is self-hosting: the K2 line ships as open weights under a Modified MIT license, so teams with their own GPU capacity can run it without paying per token at all.
Free Chinese coding models compared
| Model / Lab | Free chat interface | Free API option | License | Best for |
|---|---|---|---|---|
| DeepSeek | chat.deepseek.com | Free DeepSeek V3 / V3.1 on OpenRouter | MIT | General-purpose default, cheapest paid fallback |
| Qwen3-Coder (Alibaba) | chat.qwen.ai (Qwen-Coder mode) | qwen/qwen3-coder:free on OpenRouter | Apache 2.0 | Local/self-hosted coding, huge context |
| GLM (Zhipu/Z.ai) | chat.z.ai (no sign-up) | Self-host or check OpenRouter listings | MIT | Agentic tools: Claude Code, Cline, Roo Code, Kilo Code |
| Kimi K2 (Moonshot AI) | kimi.ai | Self-host only (no permanent free hosted API) | Modified MIT | Long, multi-step coding agent sessions |
Three ways to actually use these for free
- Free web chat. The lowest-friction option for everyday coding help: paste code or describe a bug into chat.deepseek.com, chat.qwen.ai, chat.z.ai, or kimi.ai. No API key, no install.
- Free-tier models via OpenRouter. If you want to call a model from your own script,
IDE plugin, or agent harness, OpenRouter hosts zero-cost variants of several of these models (look for
model IDs ending in
:free). Rate limits are tighter than paid tiers, but there's no card required. - Self-host the open weights. DeepSeek, Qwen, GLM, and Kimi all publish permissively licensed weights (MIT or Apache 2.0). If you have the GPU hardware, running the model yourself removes per-token cost and rate limits entirely — the only expense is your own compute.
Which one should you pick?
For most people who just want free coding help without committing to a workflow, start with DeepSeek's chat — it's the most general-purpose and has the most free-tier API coverage on OpenRouter as a fallback. If you're self-hosting or need a very large context window for a big codebase, Qwen3-Coder's Apache 2.0 license and 256K+ token context make it the more flexible pick. If you're already using an agentic coding tool like Claude Code or Cline and want to swap in a free model, GLM is the one specifically validated in that setup. And if your work involves long, multi-step agent sessions rather than single Q&A exchanges, Kimi K2 is built for exactly that, provided you're willing to self-host for the fully free path.
FAQ
Are DeepSeek, Qwen, GLM, and Kimi actually free to use for coding?
Yes, in three ways: their hosted web chat products (chat.deepseek.com, chat.qwen.ai, chat.z.ai, kimi.ai) are free to use in a browser with no card required, several of the underlying models are listed as zero-cost variants on OpenRouter, and the open weights themselves are released under permissive licenses (MIT or Apache 2.0 for most) so you can self-host them without per-token fees.
Which free Chinese AI model is best for coding right now?
There isn't one universal winner — DeepSeek is the best general-purpose default with the broadest free-tier API coverage, Qwen3-Coder is strongest for self-hosted work with very large codebases thanks to its context window, GLM is specifically tuned for agentic tools like Claude Code and Cline, and Kimi K2 is built for long, multi-step coding sessions.
Is Qwen Code still free in 2026?
Alibaba's official free OAuth quota for the Qwen Code CLI was reduced and then closed entirely by
April 2026, replaced by a paid $50/month Coding Plan. Qwen3-Coder itself is still usable for free
through Qwen Chat's web interface or via the free qwen/qwen3-coder:free listing on
OpenRouter.
Can I use these models inside Claude Code, Cursor, or Cline?
Generally yes, if the tool lets you point at a custom API endpoint or OpenRouter. GLM in particular is documented as performing well inside Claude Code, Cline, Roo Code, and Kilo Code specifically, since it's tuned for the same tool-calling patterns those harnesses use.
Are these models safe to use for private or proprietary code?
Treat any hosted third-party chat or API — Chinese or otherwise — the way you'd treat any external service: read the provider's data retention and training-use terms before pasting proprietary code into it. If that's a concern, the self-hosting option removes the question entirely, since your code never leaves your own infrastructure.
What's the catch with "free" AI coding models?
Mainly rate limits and stability. Free web chat tools and free-tier API models are typically capped on requests per day and can be slower or deprioritized under load compared to paid tiers, and free quotas have historically been reduced or removed on short notice (as happened with Qwen Code's free OAuth tier). For serious daily use, expect to eventually pay for a plan or self-host.
Can I self-host these models instead of using an API?
Yes. DeepSeek, Qwen3-Coder, GLM, and Kimi K2 are all released as open weights under permissive licenses, so anyone with sufficient GPU capacity can download and run them locally or on their own servers with no per-token cost.
Do I need a Chinese phone number or ID to sign up?
No — the international web chat products (chat.deepseek.com, chat.qwen.ai, chat.z.ai, kimi.ai) are built for a global audience and support standard email or third-party sign-up, not a Chinese phone number.
Further reading:
- Why Are AI Models So Expensive? How to Plan Model Use for Low-Cost, High-Output Work (2026 Guide)
- AI Coding Agent Pricing and Usage Limits FAQ: Claude Code vs Cursor vs Copilot vs Codex (2026)
- Top AI Models to Use in Cursor IDE for App Development (2026 Guide)
- How to Make Claude Code Faster and Choose the Right Model for Software Development (2026 Guide)