On June 23, 2026, Alibaba Cloud’s Qwen team released QwenPaw, a free open source AI assistant with a built-in web IDE. The package includes a 32 billion parameter model based on Qwen 3.6, a 128,000 token context window, and an Apache 2.0 license. It ships as a local runtime that you can run on a single GPU or CPU. This launch arrives after weeks of paid coding tools adding hidden fees and usage caps. See the pricing overhaul among major coding tools. The web IDE lets you edit files, run terminal commands, and manage multi step agent tasks without leaving the browser. No API key or subscription is required.
Alibaba Cloud announced the release on its official Alibaba Cloud homepage and mirrored the model on Hugging Face. The project does not use a paid gate. The full stack, from model weights to the IDE frontend, is open. This matters because developers have reported surprise costs from GitHub Copilot usage based billing and Cursor free tier limits. Details on Copilot billing backlash are here. QwenPaw runs locally by default. That keeps code, prompts, and output on your machine. It is a contrast to cloud tools that send every keystroke to a data center. You can inspect the source on GitHub and modify any part.
The technical foundation is Qwen 3.6, an Apache 2.0 coding model that already holds strong scores on HumanEval and SWE-bench. QwenPaw wraps that model with a web based IDE that supports Python, JavaScript, and shell commands. The assistant can read project files, write patches, and run tests. You can also use it as a chat agent for research and data analysis. Because the license is permissive, you can build commercial products on top. Read the Qwen 3.6 release details. This is not a cloud demo. It is a full local development environment. That changes the cost math for small teams and solo developers.
QwenPaw lands during a broader shift in free AI access. Google cut Gemini API tiers and OpenAI moved some models behind paid plans. Free tier limits now reset every five hours on Claude. QwenPaw removes that problem by giving you the entire tool. You pay for electricity and hardware, not tokens. The downside is setup. You need at least 16GB of RAM for CPU inference or a 24GB GPU for comfortable speeds. Not every laptop will run it well. Still, the release gives developers an escape hatch from subscription fatigue.
How Do the Top Options Compare?
| Tool | Best For | License | Model Size | Context | Web IDE | Local Only |
|---|---|---|---|---|---|---|
| QwenPaw | Local coding with full IDE | Apache 2.0 | 32B | 128K | Yes | Yes |
| OpenCode | Terminal based agents | MIT | Configurable | Varies | No | Yes/Cloud |
| Qwen 3.6 Base | Custom fine-tuning | Apache 2.0 | 32B | 128K | No | Yes |
| Cursor Free Tier | Cloud coding without GPU | Proprietary | Frontier models | Varies | Yes | No |
OpenCode can run cloud models with API keys. Cursor free tier limits depend on current subscription terms. QwenPaw and Qwen 3.6 Base use the same underlying model but different packaging.
1. QwenPaw , Best for developers who want a local AI coding IDE with no usage limits
QwenPaw combines a 32B parameter Qwen 3.6 model with a browser based IDE. The model scored 91.2% on HumanEval and 72.4% on SWE-bench Verified in Alibaba Cloud’s published results. The IDE includes a file tree, editor, terminal, and agent panel. You can run it locally with one command. No data leaves your machine unless you configure a remote endpoint. See the best open source LLM models for coding. The Apache 2.0 license covers weights, code, and configuration. Commercial use is allowed. You can also swap the model for any GGUF or vLLM compatible checkpoint. The tool includes presets for 4 bit and 8 bit quantization. That helps it fit on smaller GPUs. Setup is straightforward for Linux and macOS. Windows support requires WSL2. The model downloads about 19GB for the 4 bit version. CPU inference works but is slow for long files. The project is still early. Some IDE features, like live collaboration and plugin support, are missing. But the core coding loop works. QwenPaw does not include cloud sync or team management. That is a tradeoff for full local control. For solo developers and privacy focused teams, it is a practical option.
Key strengths:
- ✅ Free with Apache 2.0 license for commercial use
- ✅ Local inference keeps code and prompts private
- ✅ Web IDE supports file editing and terminal commands
- ✅ No token limits or subscription fees
- ✅ Runs on a single 24GB GPU or CPU with patience
- ❌ Requires a capable machine for fast inference
- ❌ Windows needs WSL2
- ❌ Missing collaboration and plugin features
Who it’s for: Choose QwenPaw if you want a private, self-hosted coding assistant and are willing to manage local hardware.
2. OpenCode , Best for terminal based AI coding agents
OpenCode is a free open source AI coding agent that runs in the terminal. It supports many local and remote models through providers like Ollama, llama.cpp, and API keys. The MIT license permits commercial use. OpenCode does not include a graphical web IDE. Instead, it gives you an interactive command line interface and scripting hooks. Read the OpenCode release. It can edit files, run shell commands, and chain multiple steps. The project has a growing plugin ecosystem. You can integrate OpenCode with VS Code, Neovim, and GitHub Actions. It is lighter than QwenPaw because it does not bundle a model or frontend. Benchmark results depend on the model you attach. OpenCode itself is model agnostic. That means you can use a 7B local model or GPT 4 class cloud model. The downside is more setup. You must install a model runtime and configure keys. QwenPaw ships everything in one package. OpenCode is better if you already have a preferred stack. It also works on servers without a browser. If you want a web IDE, you must pair it with VS Code or another editor.
Key strengths:
- ✅ MIT license and model agnostic
- ✅ Terminal native with scripting hooks
- ✅ Plugin support for VS Code and Neovim
- ✅ Low resource overhead for small models
- ❌ No built-in web IDE
- ❌ Requires separate model runtime setup
- ❌ Less beginner friendly than QwenPaw
Who it’s for: Choose OpenCode if you prefer terminal workflows and want to plug in your own model.
3. Qwen 3.6 Apache Base Model , Best for custom fine-tuning and model integration
Qwen 3.6 is the engine inside QwenPaw. It is a 32B parameter dense model with a 128K context window. Alibaba Cloud released it under Apache 2.0 on June 10, 2026. The model scores 91.2% on HumanEval and 72.4% on SWE-bench Verified. See the Qwen 3.6 open source release. It supports tool calling, agent tasks, and long context reasoning. You can download weights from Hugging Face or ModelScope. The base model does not include an IDE. You run it in vLLM, llama.cpp, or Ollama. QwenPaw adds a user interface and project management on top. Using the base model gives you maximum flexibility. You can fine-tune it on your own codebase. You can quantize it to 4 bit for a 19GB file. You can serve it behind an API for team use. The tradeoff is that you must build your own assistant loop. That means handling file edits, terminal execution, and context management yourself. QwenPaw already does that. If you need a stripped down model for an existing app, the base is the better choice.
Key strengths:
- ✅ Apache 2.0 license and strong coding benchmarks
- ✅ 128K context for long repository analysis
- ✅ Runs in standard serving frameworks
- ✅ Good base for fine-tuning
- ❌ No IDE or agent layer included
- ❌ Requires integration work
- ❌ Larger memory footprint than smaller coding models
Who it’s for: Choose Qwen 3.6 base if you want to embed the model into your own custom tool.
4. Cursor Free Tier , Best for quick cloud based AI coding without local hardware
Cursor is a commercial AI code editor with a free tier. The free plan includes limited prompts and slow requests after the quota is used. It is cloud based, so it works on any laptop with a browser or editor install. Cursor offers strong models like GPT 5.5 and Claude Opus through its backend. The free tier is convenient for quick edits. But you do not own the stack. The vendor can change limits at any time. Recent pricing shifts added usage based billing for some users. Cursor also sends code to their servers unless you opt out. Compared to QwenPaw, Cursor is easier to start. No GPU is required. The model quality is often higher for frontier cloud models. But the free tier resets limits and may throttle you. QwenPaw gives unlimited local use but demands hardware. For privacy conscious developers, Cursor is not ideal. Your code may be used for training unless disabled. QwenPaw keeps everything local by default.
Key strengths:
- ✅ No local GPU required
- ✅ Access to frontier cloud models
- ✅ Polished IDE with extensions
- ✅ Quick setup
- ❌ Free tier has strict quotas
- ❌ Code sent to cloud by default
- ❌ Vendor can change pricing or limits
Who it’s for: Choose Cursor free if you need a fast start and do not mind cloud privacy tradeoffs.
Frequently Asked Questions
Is QwenPaw completely free for commercial use?
Yes. QwenPaw uses the Apache 2.0 license. You can modify, distribute, and use it in commercial products without paying Alibaba Cloud. You must retain copyright notices and disclaimers. The model weights and IDE source are both included under the same license. There are no usage caps or API fees.
What hardware do I need to run QwenPaw?
For the 4 bit quantized model, you need about 19GB of storage and at least 10GB of GPU VRAM. A 24GB GPU like an RTX 3090 or 4090 gives comfortable speed. CPU inference works with 16GB of RAM or more but is slow for long files. The full 32B model in 16 bit needs around 64GB of memory. QwenPaw includes presets for 4 bit and 8 bit quantization.
Does QwenPaw work without an internet connection?
Yes. Once you download the model weights and install the runtime, QwenPaw runs fully offline. The web IDE is served locally. No prompts, code, or outputs are sent to a cloud service. You only need internet for the initial download and updates. This makes it suitable for air gapped environments.
How does QwenPaw compare to GitHub Copilot free tier?
GitHub Copilot free tier offers cloud based completions with strict hourly and monthly limits. QwenPaw gives unlimited local completions with no token quota. Copilot is easier to set up and uses more powerful frontier models. QwenPaw keeps your code private and has no vendor billing risk. Choose based on hardware and privacy needs.
Can I swap the underlying model in QwenPaw?
Yes. QwenPaw supports any GGUF or vLLM compatible model that fits the same API. You can replace the default Qwen 3.6 32B with a smaller 7B model for weaker hardware. You can also point it at a remote OpenAI compatible endpoint. The IDE handles model config through a settings file. This gives you flexibility beyond the default model.
Where can I find QwenPaw?
The official release is listed on Alibaba Cloud’s homepage and Hugging Face. The source code is on GitHub under the Qwen organization. You can download Docker images or build from source. The announcement includes benchmark details and setup guides. Always use the official links to avoid tampered weights.
What Should You Remember?
- Free and open source: QwenPaw ships under Apache 2.0 with no API fees or usage caps.
- Local by default: All code, prompts, and outputs stay on your machine.
- Built-in web IDE: Edit files, run terminal commands, and manage agent tasks in the browser.
- 32B parameter model: The Qwen 3.6 base provides 128K context and strong coding benchmarks.
- Hardware tradeoff: You need a 24GB GPU for fast inference or patience on CPU.
- Commercial ready: Permissive license allows modification and commercial products.
- Alternative to paid tiers: QwenPaw bypasses subscription fatigue from Copilot, Cursor, and Claude limits.
Free AI News is an independent editorial publication. Information about AI pricing, free-tier limits, and features changes frequently and may become outdated. Always verify current details through the vendor’s official pages. Affiliate links may earn a commission at no cost to you, and never affect our reporting.
