Almanac
Guide · Beginner

Alibaba's Qwen: The Open-Weight AI Lab Taking on the World's Frontier Models

AlibabaBeginneractive·v1 · live·generated 5d ago
TL;DRAlibaba's Qwen team has quietly built one of the most prolific AI model families in the world, releasing dozens of open-weight models across text, code, vision, audio, and math — and making most of them free to download and use. Starting from a single language model, Qwen has grown into a full-stack AI platform that competes directly with OpenAI and Anthropic, while its open-source releases have reshaped what developers outside big tech can build.

Key takeaways

  • The Qwen3-Coder flagship is a 480-billion-parameter model that claims performance comparable to Anthropic's Claude Sonnet 4 on coding tasks.
  • Qwen2.5 was described as potentially the largest open-source model release in history, spanning seven sizes from 0.5B to 72B parameters.
  • The Qwen3.5 vision-language flagship outperforms GPT-5.2, Claude 4.5 Opus, and Gemini-3 Pro on 28 of 44 vision benchmarks — with open weights under Apache 2.0.
  • Qwen3.7-Max, the closed-weights flagship, ranks fifth on the Artificial Analysis Intelligence Index and claims the lowest hallucination rate among frontier models tested.
  • A research study found Qwen 2.5 showed the most extreme geopolitical bias shift of any lab tested — an 18x increase in China-favourability after post-training.
  • Alibaba has recently shifted its top-tier models from open to closed weights, signaling a more revenue-focused strategy.

What Alibaba's Qwen is

Alibaba is one of China's largest technology companies, and its Qwen team is its AI research and model division. Since 2023, Qwen has released a large and growing family of AI models — tools that can read, write, reason, write code, understand images, transcribe audio, and analyze video. If you've heard of ChatGPT or Claude, Qwen is Alibaba's equivalent: a suite of AI assistants and developer APIs, with one major twist — most of the models are free to download and run yourself.

Why it matters

For most of AI's recent history, the most powerful models were locked behind paywalls and APIs controlled by a handful of U.S. companies. Qwen changed that calculus. When Alibaba released Qwen2.5 — described at the time as potentially the largest open-source model release in history — it put seven different model sizes (from tiny mobile-friendly versions to 72-billion-parameter workhorses) into the hands of any developer, for free. That matters because it means startups, researchers, and companies in any country can build on frontier-quality AI without paying per-query fees or agreeing to usage restrictions.

What Qwen can do

Qwen isn't one model — it's a whole ecosystem. Here's a plain-language map of the main branches:

  • Language models: The core Qwen series handles reading, writing, summarizing, and reasoning in dozens of languages. The Qwen-MT Turbo translation model supports 92 languages covering over 95% of the world's population.
  • Coding: Qwen3-Coder, the flagship coding model, is a 480-billion-parameter system that claims performance on par with Anthropic's Claude Sonnet 4 on programming tasks. It can handle files up to 1 million words long.
  • Vision: The Qwen3.5 vision-language family can look at images and answer questions about them. The flagship version outperformed GPT-5.2, Claude 4.5 Opus, and Gemini-3 Pro on 28 out of 44 vision benchmarks.
  • Audio and video: Qwen2.5-Omni can process text, images, audio, and video all at once, responding in real time with both text and synthesized speech. Qwen2-VL can understand videos longer than 20 minutes.
  • Math and reasoning: QwQ-32B is a reasoning-focused model trained with reinforcement learning — a technique that teaches the model to check its own work — aimed at mathematics and complex problem-solving.

How it's distributed

Most Qwen models are available on Hugging Face and ModelScope (two popular platforms for sharing AI models), as well as GitHub. Hosted versions run on Alibaba Cloud. This wide distribution means developers can pick the version that fits their budget and hardware — from a tiny model that runs on a phone to a massive one that needs a server.

Recent developments

Alibaba has been releasing new models at a rapid pace. In early 2026, the Qwen3.5 family arrived with strong vision capabilities and small models that punch well above their weight — the 9-billion-parameter version outperformed OpenAI's 120-billion-parameter model on several tests. Shortly after, Qwen3.7-Max launched as a closed-weights flagship targeting long-running "agentic" tasks — jobs where the AI works autonomously over many steps, like writing and debugging an entire software project.

That shift from open to closed weights for the top-tier models is notable. It suggests Alibaba is moving toward a more commercial strategy for its most powerful systems, while continuing to release smaller open models for the developer community.

Things to keep in mind

A published research study found that Qwen 2.5 showed the most extreme geopolitical bias of any lab tested — its post-training process increased China-favourability by 18 times compared to the base model. The researchers argue this kind of bias is introduced during the fine-tuning process, not baked into the raw training data. It's a reminder that open weights don't automatically mean neutral or unbiased, and that anyone deploying these models in sensitive contexts should evaluate them carefully.

Where it's heading

Qwen's trajectory points toward two things happening at once: continued open-weight releases for developers and researchers, and a growing suite of closed, commercially-focused flagship models for enterprise customers. The breadth of the model family — covering more modalities and languages than almost any other lab — suggests Alibaba is building toward a general-purpose AI platform, not just a chatbot.

The Qwen model family at a glance

Timeline

  1. Qwen-VL-Plus and Qwen-VL-Max launched — first major multimodal upgrade

  2. Qwen1.5-110B released — first Qwen model over 100B parameters

  3. Qwen2.5 launches — described as potentially the largest open-source model release in history

  4. Qwen2.5-Turbo extends context to 1M tokens

  5. QwQ-32B released — Qwen's first RL-scaled reasoning model

  6. Qwen3.5 vision-language family released — tops 28 of 44 vision benchmarks

  7. Qwen3.7-Max launched as closed-weights agentic flagship

  8. Qwen3-Coder-480B released — claims parity with Claude Sonnet 4 on coding

Related topics

QwenQwen2.5Qwen3Qwen2.5-MaxQwen TeamMixture of ExpertsHugging FaceModelScopeOpenAIAnthropic

FAQ

What is Qwen, exactly?

Qwen is the name of Alibaba's family of AI models — covering text, code, images, audio, and video — as well as the team that builds them. Think of it as Alibaba's answer to OpenAI's GPT or Google's Gemini.

Are Qwen models free to use?

Most Qwen models are released as open weights under the Apache 2.0 license, meaning you can download and use them freely. However, Alibaba's newest top-tier models (like Qwen3.7-Max) are now closed-weights and available only via API.

How does Qwen compare to ChatGPT or Claude?

Qwen's flagship models rank in the top five to seven on major AI leaderboards, trailing the very best from OpenAI, Anthropic, and Google — but the open-weight versions are among the most capable models anyone can freely download and run.

Should I be concerned about bias in Qwen models?

A published study found that Qwen 2.5 showed the largest geopolitical bias shift of any lab tested — an 18x increase in China-favourability after post-training — so users deploying Qwen in sensitive contexts should be aware of this finding.

What is Qwen good at?

Qwen has specialized models for coding, math, translation (92 languages), vision, audio, and long-document reasoning — making it one of the broadest AI model families available, especially for developers who want to run models locally.

Stay current

Call Me Almanac pairs the week's AI news with guides like this one — Midweek & Sunday.

Versions

  • v1live5d ago

Related guides (4)

More on Alibaba (6)

7The Batch·18d ago·source ↗

Alibaba releases Qwen3.5 open-weights vision-language model family with MoE architecture across eight sizes

Alibaba released the Qwen3.5 family of eight open-weights vision-language models ranging from 0.8B to 397B parameters, built on a mixture-of-experts architecture with mixed attention and Gated DeltaNet layers. The flagship Qwen3.5-397B-A17B outperforms GPT-5.2, Claude 4.5 Opus, and Gemini-3 Pro on 28 of 44 vision benchmarks, while the 9B model surpasses OpenAI's gpt-oss-120B on most language tasks. Open weights are available under Apache 2.0, with hosted agentic variants (Qwen3.5-Plus, Qwen3.5-Flash) available via Alibaba Cloud. The release is notable for strong small-model efficiency and comes amid reported team departures following the Qwen3 rollout.

6The Batch·15d ago·source ↗

Alibaba's Qwen3.7-Max positions as top Chinese LLM with closed weights and agentic focus

Alibaba released Qwen3.7-Max, a closed-weights proprietary model targeting long-running agentic tasks like coding and scientific discovery, with a 1M-token context window and 208 tokens/second output speed. The model ranks fifth to seventh on the Artificial Analysis Intelligence Index, trailing leading U.S. models from OpenAI, Anthropic, and Google but claiming the lowest hallucination rate among frontier models tested—partly by declining to answer over half of prompts. Alibaba's training approach separates task, agentic harness, and verifier components to prevent overfitting to specific setups. The release continues Alibaba's strategic shift from open to closed weights for top-tier models, with leadership changes in the Qwen team suggesting a revenue-focused pivot.

7Qwen Research·1mo ago·source ↗

Qwen2.5-Turbo Extends Context Length to 1M Tokens

Alibaba's Qwen team has released Qwen2.5-Turbo, extending the model's context window from 128K to 1 million tokens (approximately 1 million English words). The update includes optimizations for both model capabilities and inference performance at extreme context lengths. The model is available via API and through HuggingFace and ModelScope demos.

8Qwen Research·1mo ago·source ↗

Qwen2.5-LLM: Alibaba releases open-weight language models from 0.5B to 72B

Alibaba's Qwen team releases the Qwen2.5 series of decoder-only dense language models, open-sourcing seven variants spanning 0.5B to 72B parameters. The release targets production use cases in the 10-30B range and mobile deployments at 3B scale. This represents a significant expansion of the open-weights frontier from a Tier 1 Chinese AI lab.

7Qwen Research·1mo ago·source ↗

Qwen2-VL: Alibaba Releases Latest Vision-Language Model with Extended Video Understanding

Alibaba's Qwen team has released Qwen2-VL, the latest iteration of their vision-language model series built on the Qwen2 foundation. The model claims state-of-the-art performance on visual understanding benchmarks including MathVista, DocVQA, RealWorldQA, and MTVQA. A notable capability is understanding videos exceeding 20 minutes in length for question answering, dialog, and content creation tasks.

8Qwen Research·1mo ago·source ↗

Qwen2.5: Large-Scale Open-Source Foundation Model Family Release

Alibaba's Qwen team has released Qwen2.5, described as potentially the largest open-source model release in history, following three months of development after Qwen2. The release encompasses a family of foundation models with improvements in knowledge and reasoning capabilities. The announcement targets developers who have been building on Qwen2 and incorporates feedback from that community.