Qwen

Qwen (通义千问) is Alibaba Cloud's open-weight large language model family, developed by the Qwen team and released under the Apache 2.0 licence. Since its first public release in 2023, Qwen has grown into one of the most comprehensive open-weight LLM ecosystems, spanning general-purpose language models, code specialists, vision-language models, and dedicated reasoning models.

Model families

• Qwen3 (April 2025): The flagship generation with dense models at 0.6B, 1.7B, 4B, 8B, 14B, and 32B parameters, plus MoE variants at 30B-A3B and 235B-A22B. Dense models support up to 128K token context; MoE variants extend to 256K+ and 1M tokens with extrapolation. Qwen3 introduces a built-in thinking mode for step-by-step reasoning that can be toggled on or off per request.
• Qwen3.5 (February 2026): A 397B-A17B MoE model extending the series, supporting up to 32K tokens in most configurations.
• Qwen3.6 (April 2026): Dense 27B and MoE 35B-A3B models with a 262K native context window extensible to 1 million tokens. Qwen3.6-Max-Preview is a proprietary hosted frontier model (parameter count undisclosed) optimised for agentic coding and long-context reasoning.
• Qwen3-Coder: The coding-specialist branch of Qwen3. Qwen3-Coder-480B-A35B-Instruct (480B total / 35B active) supports 256K native context and 1M with extrapolation. Qwen3-Coder-Next (80B total / 3B active) targets efficient on-device and agent use, achieving performance rivalling models with 10–20× more active parameters.
• Qwen2.5-VL (January 2025): Vision-language models at 3B, 7B, 32B, and 72B parameters. Capable of understanding images, charts, documents, and videos over 1 hour long; can act as a visual agent performing computer and phone use tasks.
• Qwen3-VL: The next-generation multimodal branch of the Qwen3 series, with a 32B-Instruct variant on Hugging Face.
• QwQ: Reasoning-specialist models designed for complex mathematical and logical problems. QwQ-32B is competitive with DeepSeek-R1 and o1-mini on hard reasoning benchmarks.

Access methods

Self-hosting: All open-weight checkpoints are published on Hugging Face under the QwenLM organisation. Models run locally or on your own infrastructure via Ollama (ollama run qwen3:30b-a3b), vLLM (≥0.8.4), LM Studio, llama.cpp, MLX, and KTransformers. Smaller models (0.6B–8B) fit on consumer GPUs.

Alibaba Cloud API (DashScope / Model Studio): Access any Qwen variant through an OpenAI-compatible endpoint at dashscope.aliyuncs.com. Drop-in replacement for the OpenAI SDK — just swap the base_url and set DASHSCOPE_API_KEY. Regional endpoints available in China, Singapore, US (Virginia), and Hong Kong. API tiers include qwen-turbo (low cost), qwen-plus (balanced), and qwen-max (highest quality).

Third-party providers: Qwen models are widely available on Together AI, Fireworks AI, OpenRouter, and other aggregator APIs, typically with competitive per-token pricing.