Alibaba

corporate

china · Founded 1999 · Part of Alibaba Group · IPO 2014 (NYSE: BABA) · Market cap: $299B (HKG: 9988 / NYSE: BABA) as of 2026-03

The Qwen team is Alibaba Group's flagship large model effort and the most prolific contributor to the open-source AI ecosystem globally, with over 1 billion total downloads on HuggingFace. Originally developed within the DAMO Academy (established 2017) under the Tongyi Qianwen brand, the team evolved through predecessor projects including M6 (10T sparse MoE, 2021) and OFA (ICML 2022). Models are served commercially through Alibaba Cloud Model Studio.

The Qwen series has progressed from Qwen (2023) through Qwen3.5 (February 2026) at remarkable speed, spanning dense and MoE architectures across text, code, math, vision, audio, and omni modalities. Qwen3 introduced hybrid thinking with seamless reasoning mode-switching; Qwen3.5 (397B-A17B) supports 201 languages and 1M-token context. Alibaba has committed $53 billion over three years (2025–2028) for AI and cloud infrastructure, the largest private technology investment in Chinese history.

In March 2026, the team underwent a significant restructuring: Tech Lead Junyang Lin, post-training head Bowen Yu, and coding lead Binyuan Hui all departed. Alibaba responded by creating a new "Token Hub" business group under CEO Eddie Wu, with CTO Jingren Zhou (ACM Fellow, PhD Columbia, ex-Microsoft Research) taking direct oversight of Qwen model development. Former Google DeepMind researcher Hao Zhou was recruited to lead post-training RL.

Website Wikipedia HuggingFace GitHub
frontieropen-weightmultimodalmoe

People

News

Outputs (30)

Think-Anywhere

paper

<p>Introduces <strong>Think-Anywhere</strong>, enabling LLMs to invoke reasoning blocks at arbitrary token positions during code generation, rather th…

Qwen3.5 5

model

Fourth generation of the Qwen model family. Qwen3.5-397B-A17B native VLM built on Qwen3-Next architecture. 201 languages and dialects supported. 1M co…

★ 2.2k

RationaleRM

dataset

Dataset for training reasoning reward models.

DeepPlanning

dataset

Benchmark for evaluating complex agentic planning.

Qwen3Guard

model

Safety-specialized guardrail model for real-time token stream moderation.

★ 439

Qwen-Image 2

model

Image generation model family. Qwen-Image (20B MMDiT) excels at text rendering in Chinese and English. Qwen-Image-2.0 (7B) unifies generation and edit…

★ 7.6k ↓ 223.7k

Qwen-Code

library

CLI tool for terminal-based agentic coding and codebase analysis.

★ 20.9k

Qwen3-Coder 2

model

Third-generation coding-specialized model family including Qwen3-Coder-480B and Qwen3-Coder-Next.

★ 16.1k ↓ 1.3M

Gated Attention for Large Language Models

paper

Awarded **Best Paper at NeurIPS 2025**, this research introduces a foundational architectural modification for transformers: applying a head-specifi…

Qwen3 9

model paper

Third generation of the Qwen model family. Introduced a massive 36-trillion token pretraining dataset. Includes dense, MoE, multimodal, speech, and re…

★ 27k 📄 40

PolyMath

dataset

High-quality mathematical reasoning dataset.

★ 43

Qwen2.5-Omni-7B

model

End-to-end multimodal model processing text, images, audio, and video with real-time speech generation. Thinker-Talker architecture. Over 80k download…

★ 4k ↓ 466.1k

QwQ-32B

model

Reasoning model following the chain-of-thought (CoT) paradigm.

↓ 54.9k

Qwen2.5-VL

model

Vision-language model with dynamic resolution processing and absolute time encoding.

QVQ

model

Multimodal reasoning model following the chain-of-thought (CoT) paradigm.

ProcessBench

dataset

Benchmark for evaluating the step-by-step logic of LLMs.

📄 1

Qwen2.5 3

model paper

The "18-trillion token" series. Released sizes: 0.5B, 1.5B, 3B, 7B, 14B, 32B, and 72B, plus a proprietary MoE model (Qwen2.5-Max).

★ 27k 📄 57

Qwen2-VL

model

Naive Dynamic Resolution for variable-resolution visual processing. Available in 2B, 8B, and 72B variants.

↓ 48.7k 📄 246

Qwen2.5-Coder 2

model paper

Coding-specialized model family. Code-specific models (0.5B-32B) pre-trained on 5.5T tokens.

★ 16.1k 📄 30

Qwen2.5-Math 2

model paper

Math-specialized model family. Includes the Qwen2.5-Math series (1.5B, 7B, 72B) which introduced Tool-Integrated Reasoning (TIR).

★ 1.1k 📄 10

Qwen2 2

model paper

Significantly improved multilingual capabilities. Sizes: 0.5B, 1.5B, 7B, 72B, and the 57B-A14B MoE.

★ 27k 📄 43

Qwen1.5

model

Major update (the "Qwen2 Beta") introducing 0.5B, 1.8B, 4B, 7B, 14B, 32B, 72B, and 110B dense models, alongside a 14B MoE variant.

Qwen2-Audio 2

model paper

Audio model family. Evolution from Qwen-Audio (2023) to Qwen2-Audio-7B (Aug 2024).

★ 2.1k ↓ 2.3k 📄 25

Qwen-Agent

library

Python framework for building agents with tool-calling, RAG, and code execution.

★ 15.7k

qwen.cpp

library

C++ implementation for efficient local inference of Qwen models.

★ 621

Qwen-VL

model

First-generation vision-language model in the Qwen family.

Qwen 3

model paper

First generation of the Qwen large language model series, spanning 1.8B to 72B parameters.

★ 20.8k ↓ 113.2k 📄 80

Chinese CLIP

model

Contrastive vision-language pretraining specifically for Chinese. Predates the Qwen branding.

★ 5.8k 📄 51

OFA: One For All

model

Unified sequence-to-sequence framework for multimodal tasks. Predates the Qwen branding.

★ 2.6k 📄 258

M6 Series

model

Multi-modality-to-multi-modality mega-transformer, including the M6-10T (10 trillion parameters) and M6-T sparse models. Predates the Qwen branding.

📄 48