Qwen3.8 🧠 Model Open source

What it is

Qwen is Alibaba's large language model family. The current generation is Qwen3.8: the hosted Qwen3.8-Max (2 August 2026), open weights for Qwen3.8-2.4T-A95B (12 August 2026, the first open release of a Max-class Qwen model), the dense Qwen3.8-27B (mid August) and Qwen3.8-Flash (26 August 2026). Earlier 2026 releases were Qwen3.5 (February) and Qwen3.6 (April). Qwen 4 is described as in training with no date.

Key features

• Apache 2.0 weights from small to 2.4T parameters, published on Hugging Face and ModelScope
• 262K-token context in the documented deployments; Qwen3.8-Flash supports a million tokens
• Native vision-language input (image and video) on Qwen3.8-27B
• Support for 201 languages and dialects
• Runs locally with llama.cpp, vLLM and SGLang

Where you can use it

In Alibaba's Qwen Studio web chat, on the QwenCloud / Alibaba Cloud Model Studio API (OpenAI- and Anthropic-compatible), and self-hosted from the open weights. We found no app in our directory that documents Qwen as a built-in model choice, so the app list is empty; developers usually reach it through the API, OpenRouter or local runners.

Pricing and rights

The open models are Apache 2.0, which permits commercial use and fine-tuning. Qwen Studio chat is free. Cloud API prices vary by model and region on Alibaba Cloud Model Studio; we did not find a single current price sheet to quote, so check the console before budgeting.

Who it is for

Developers and multilingual publishers who want strong open weights under a permissive licence, especially for Chinese and other Asian languages.

Verdict

Qwen3.8 is one of the most capable openly licensed model families you can run yourself. It is tuned for coding and agents, the big models need serious hardware, and mainstream writing apps rarely offer it as a menu option.