Alibaba Cloud has built one of the largest and most comprehensive open source AI ecosystems in the world. Since its initial release in 2023, Qwen has grown from a single language model into a full stack of open weight models spanning text generation, coding, vision, audio, speech, image generation, and real-time multimodal interaction.

The latest flagship — Qwen3.5 released February 2026 — is a 397B Mixture-of-Experts model with 17B active parameters, a 262K native context window extensible to 1M tokens, and native support for 201 languages. It competes directly with GPT-5.2 and Claude Opus 4.5 across reasoning, coding, and multimodal tasks — while remaining open weight under the Apache 2.0 license.

The result is frontier-level performance at zero per-token cost — download the weights, run locally, fine-tune on your own data, and deploy without restriction for research and most commercial uses.

Qwen's open weight models and free access options open up a range of possibilities for anyone working in AI and ML:

Run frontier-level inference locally — download Qwen3.5 or QwQ and run inference on your own hardware with no API key, no per-token costs, and no data leaving your machine. The same performance as leading closed models, on hardware you control.

Fine-tune on your own data — all open weight Qwen models can be fine-tuned using LoRA, QLoRA, and full fine-tuning workflows. Adapt a frontier model to your specific domain, dataset, or task without the cost of training from scratch.

Build multimodal applications — Qwen3.8-Flash handles text, image, video, and code input in a single model. Build applications that see, read, and reason across modalities without stitching together multiple specialized models.

Build agentic systems — Qwen3-Coder and Qwen3.8-Flash are specifically designed for agentic workflows — autonomous code generation, multi-step reasoning, and tool use. Build coding assistants, research agents, and automated pipelines on open weights.

Research multilingual AI — 201 languages natively supported. The strongest multilingual coverage of any open source model family — directly relevant for research on multilingual NLP, cross-lingual transfer, and low-resource language modeling.

Qwen's model ecosystem covers every use case and hardware profile:

Qwen3.5 — 397B MoE flagship, 17B active parameters, 1M context, 201 languages. Apache 2.0. The strongest open weight general-purpose model available.
🔗 https://huggingface.co/Qwen/Qwen3.5

Qwen3.6 — 35B-A3B dense model under Apache 2.0. April 2026. Strong performance on coding and reasoning tasks with lower hardware requirements than the 397B flagship.

Qwen3.7 — Max and Plus variants released May 2026. Extended context and improved reasoning over Qwen3.6.

Qwen3.8-Flash — latest multimodal model with 1M context window. Handles text, image, video, and code input. Strong on coding assistance, agentic workflows, and visual understanding.

Qwen3.8-Flash-Next — August 26, 2026. Open source preview of the upcoming Qwen4 architecture. Released ahead of the flagship to enable community testing and adoption. Aimed at cost efficiency.

QwQ — Qwen's reasoning model, competing directly with OpenAI o1 on complex reasoning and math tasks.

Qwen3-Coder — specialized coding model for agentic coding tasks. Strong on autonomous code generation, debugging, and multi-file editing.

All models available at huggingface.co/Qwen and github.com/QwenLM.

🔗 https://huggingface.co/Qwen
🔗 https://github.com/QwenLM/Qwen

Every major open weight Qwen model can be run three ways:

Locally via Ollama — install Ollama at ollama.ai and run ollama pull qwen3.5 to download and run Qwen3.5 locally. No API key. No per-token costs. Works on Mac, Windows, and Linux. Smaller variants like Qwen3.6 8B run on laptops with 16GB RAM.

Locally via LM Studio — download LM Studio at lmstudio.ai for a GUI-based local model runner. Search for any Qwen model and run it with one click.

Via API — access Qwen models through DashScope (Alibaba's official API), OpenRouter, Together AI, Groq, and Fireworks AI. Qwen3.5 is available on OpenRouter at $0.30/M input tokens — competitive with leading closed models.

Free via Qwen Chat — access the latest Qwen models for free at chat.qwen.ai with no API key required. The web interface includes text, image, and document input.

🔗 https://ollama.ai
🔗 https://lmstudio.ai
🔗 https://chat.qwen.ai
🔗 https://openrouter.ai/qwen

Before you download any model or write any code, paste this into Qwen Chat at chat.qwen.ai to understand which Qwen model fits your use case:

I want to build a project using Qwen open source models. Help me figure out where to start by asking one question at a time. Start with: what is the specific problem I want to solve or capability I want to build? Then ask: am I more interested in using a pre-trained Qwen model for inference, fine-tuning one on my own data, or building an agentic application? Then ask: what hardware do I have available — a laptop, a university compute cluster, or cloud credits? Then ask: what is my current experience level — have I run inference on an open source model before, fine-tuned one, or am I starting from scratch? After those four questions, give me a specific starting point — which Qwen variant fits my situation, what the first three steps are, and what resources I should read first.

Before you download a single model or write a single line of code, you will know which Qwen variant fits your use case, what your hardware requirements are, and what your first three concrete steps are.

Qwen is the most downloaded open-weight model family in the world — the foundation model more developers are choosing to fine-tune and build on than any other. 460+ open-sourced models, Apache 2.0 licensing, and frontier-level performance.

The Open Source AI Ecosystem Built to Compete With the Best in the World

Alibaba's platforms generate some of the richest ecommerce behavior data in the world. A public slice of that data is available right now on Kaggle - user clicks, cart adds, purchases, and favorites across product categories. In this challenge you use that dataset to build a user affinity profiling pipeline. Qwen Chat is your thinking partner throughout - open it in a browser tab next to your Kaggle notebook and use it to reason through what you are seeing in the data, debug when you get stuck, and interpret your results in plain language.

This is not a tutorial. There is no step-by-step guide to follow. You bring the analytical instincts. Qwen helps you think. The combination is closer to how practitioners actually work than any classroom exercise.

Go to Kaggle and search for "Alibaba User Behavior Dataset". Fork the notebook environment so you have a live Python kernel ready. The dataset contains user IDs, item IDs, category IDs, behavior types (click, purchase, cart, favorite), and timestamps. Take five minutes to load it and look at the shape before you open Qwen Chat. Know what you are working with before you ask for help thinking about it.

Open chat.qwen.ai in a tab next to your Kaggle notebook. Paste this prompt to begin:

I am working with the Alibaba User Behavior Dataset on Kaggle. It contains user IDs, item category IDs, and behavior types: click, purchase, cart add, and favorite. I want to build a user affinity profiling pipeline that assigns each user a natural language profile describing what they care about based on their behavior history. Ask me one question at a time to help me think through this. Start by asking me what I have already observed about the data.

A completed Kaggle notebook with three outputs. First, a behavior feature matrix: for each user, a weighted score across item categories reflecting how strongly they clicked, carted, and purchased in each one. Second, a cluster analysis: group users by their feature vectors using k-means or a method of your choice, and evaluate whether the clusters are meaningful. Third, a plain language profile for each cluster generated with Qwen Chat: paste your cluster summary into Qwen and ask it to describe who this user segment is and what they are likely shopping for. That last section is the one that makes the notebook portfolio-worthy - a human-readable interpretation of machine-generated segments, produced in collaboration with the same AI ecosystem that powers Alibaba's actual personalization infrastructure.

Qwen is Alibaba's open-weight AI model family - use it as your thinking partner to build a user affinity profiling pipeline on actual Alibaba ecommerce behavior data, right inside a Kaggle notebook.

Profile Real Users From Real Alibaba Data

Qwen ran a global online hackathon in 2026 across five agent-focused tracks — MemoryAgent, EdgeAgent, Autopilot Agent, AI Showrunner, and an open track. Over $70,000 in cash and credits across tracks. Every participant received free Qwen Cloud compute credits.

Future hackathons and competitions are announced at qwencloud-hackathon.devpost.com. If you are building agentic systems, RAG pipelines, or multimodal applications on Qwen — this is where to look for the next opportunity to compete and get funded.

🔗 https://qwencloud-hackathon.devpost.com

We are looking for researchers and developers to represent Qwen in their communities. Ambassadors run local events, create technical content, and get direct access to the Qwen model and product teams. Selected Ambassadors receive $100 in monthly cloud credits and a community badge. If you are active in your AI research community and want to go deeper with open-weight models, apply at qwen.ai/ambassador.

We are hiring researchers, engineers, and product builders across AI infrastructure, model development, and applied research. Open roles are listed at https://careers.alibabacloud.com/en/home?lang=en. 

If the work you do in this challenge points toward a research direction you want to pursue professionally, we want to hear from you.

Qwen is one of the most downloaded open source model families in the world - over 1 billion downloads, Apache 2.0 licensed, and built for the community to use, fine-tune, and build on. 

Whether you want to contribute, participate in events, or connect with the community, there are many ways to get involved.

Qwen

The Open Source AI Ecosystem Built to Compete With the Best in the World

Profile Real Users From Real Alibaba Data

Opportunities to Learn and Work With Us