Qwen in 2026: Alibaba's LLM Family Explained

Qwen (pronounced "chwen") is Alibaba Cloud's family of large language models, released as open-source under various licenses. In 2026, Qwen has become one of the strongest open-source LLM families — competitive with Llama 3 and often outperforming it on specific tasks.

The Qwen model family

Why Qwen stands out

Qwen2.5-72B consistently ranks near the top of open-source LLM leaderboards. Key strengths:

• Long context — native 128K token context window across the 72B model
• Multilingual — strong performance in Chinese, Japanese, Korean, and European languages, not just English
• Coding — Qwen2.5-Coder-32B is one of the best local coding models available
• Math — Qwen2.5-Math beats most models its size on math benchmarks
• Apache 2.0 — commercial use allowed without restrictions

Qwen vs Llama 3 vs Mistral

How to run Qwen locally

# Via Ollama
ollama pull qwen2.5          # 7B default
ollama pull qwen2.5:72b      # full 72B (requires ~45GB RAM)
ollama pull qwen2.5-coder    # 7B coding variant
ollama pull qwen2.5-coder:32b  # best local coding model

ollama run qwen2.5

Qwen via API

Alibaba Cloud offers Qwen models through its DashScope API. Pricing is competitive with DeepSeek — significantly cheaper than OpenAI or Anthropic for comparable capability:

• Qwen2.5-72B: ~$0.40/1M input tokens, ~$1.20/1M output tokens
• Qwen2.5-7B: ~$0.05/1M input tokens, ~$0.10/1M output tokens

Qwen 3.8 — what's new

Qwen 3.8 is Alibaba's latest release in the Qwen family, launched in mid-2026. It introduces a hybrid architecture that combines dense attention layers with mixture-of-experts (MoE) blocks — getting more capability per active parameter than the pure-dense Qwen2.5 series.

Key improvements in Qwen 3.8:

• Thinking mode — Qwen 3.8 supports a toggleable chain-of-thought reasoning mode similar to o1, producing significantly better results on math and logic tasks when enabled
• Hybrid MoE architecture — activates only a subset of parameters per token, making inference faster and cheaper than comparably-sized dense models
• Stronger coding — notable improvement on HumanEval vs Qwen2.5 at equivalent sizes
• 128K context — maintained from the 2.5 series
• Apache 2.0 — same permissive license as Qwen2.5

ollama pull qwen3.8        # latest Qwen 3.8 default
ollama pull qwen3.8:32b    # 32B variant
ollama run qwen3.8

For most practical tasks Qwen 3.8 outperforms Qwen2.5 at the same parameter count. If you're already running Qwen2.5 locally, upgrading is worth benchmarking on your specific workload before committing.