# What is Qwen3.5, and how does Alibaba's open model see, think and act?

Date: 2026-02-17 · Author: Robert Scales · https://guidetoartificialintelligence.com/en/learn/what-is-qwen3-5/

**Answer:** Qwen3.5, released by Alibaba on February 16, 2026, is a new family of AI models whose first open-weight member, Qwen3.5-397B-A17B, reads text, images and video, supports 201 languages and is free to download under the Apache 2.0 license. It uses only 17 billion of its 397 billion parameters for each token, and Alibaba says it is 60% cheaper to run than its predecessor and built for agents that act inside apps.

Yesterday, Monday, February 16, 2026, Alibaba's Qwen team released **Qwen3.5**, the next generation of the model family it first made public in 2023. The first model in the series is **Qwen3.5-397B-A17B**, and its weights, the trained numbers that make up the model, are free to download from Hugging Face under the permissive **Apache 2.0** license. That means companies can run it on their own servers, adapt it, and use it commercially.

Alibaba presents it as a model "built for the agentic AI era," able to carry out complex tasks on its own. According to the wire report carried by TimesLIVE, the company says it is 60% cheaper to use than the previous generation and eight times better at handling large workloads.

## What Alibaba released

The name describes the model's size: 397 billion parameters in total, of which 17 billion ("A17B," for active) are used for any single token. Here are the key facts from its model card:

| Feature | Qwen3.5-397B-A17B |
|---|---|
| Total / active parameters | 397 billion / 17 billion |
| Inputs | Text, images and video |
| Languages and dialects | 201 |
| Context window | About 262,000 tokens, extendable to about 1 million |
| License | Apache 2.0 (open weights) |
| Default behavior | Thinks before answering; can be switched off |

There is also **Qwen3.5-Plus**, a hosted version on Alibaba Cloud Model Studio with a 1-million-token context by default and built-in tools it can choose to use. You can also talk to the model in the Qwen Chat web app.

On benchmarks Alibaba published, Qwen3.5 lands close to the leading American models without beating them across the board. On SWE-bench Verified, which checks whether a model can fix real software bugs, it scores 76.4, against 80.0 for GPT-5.2, 80.9 for Claude Opus 4.5 and 76.2 for Gemini 3 Pro. The wire report notes that Alibaba's comparisons leave out DeepSeek, its main Chinese rival.

## How it works

Three design choices explain the efficiency and abilities.

**Mixture of experts.** Instead of one giant network that runs in full for every word, the model holds 512 smaller "expert" sub-networks. For each token, a router picks 10 of them, plus one shared expert that always runs. A hospital is a useful picture: hundreds of specialists on staff, but each patient sees only the few who are relevant. You get the knowledge of a very large model at the running cost of a much smaller one.

**Hybrid attention.** **Attention** is how a model relates each word to the others before it. The standard method gets expensive as text grows, because every word is compared with every earlier word. Qwen3.5 mixes in a cheaper method called **Gated DeltaNet**, a form of linear attention that keeps a running summary instead of re-reading everything. Its 60 layers follow a pattern of three fast DeltaNet layers, then one full attention layer. This is a large part of why long documents become affordable.

**Vision from the start.** Many models add image understanding afterward, as a separate add-on. Qwen3.5 was trained with **early fusion**: text, image and video tokens were mixed during training itself. The Qwen team says this one model matches its previous text models and beats its earlier vision-language models. It also underpins what Alibaba calls **visual agentic capabilities**, meaning the model can look at a phone or desktop screen and take actions in apps on its own.

## Why it matters

Qwen3.5 strengthens a pattern seen over the past year: some of the most capable models anyone can download come from Chinese labs, and they are closing the gap with closed American models on many tests. For businesses, an open-weight model that runs cheaply and understands 201 languages and dialects is a practical option, especially where data must stay on their own servers.

There is also a domestic race. The release came two days after ByteDance launched Doubao 2.0, the upgrade to China's most-used chatbot app, which was also pitched for AI agents. Alibaba is trying to grow its own Qwen chatbot, and the wire report notes that DeepSeek is widely expected to release a new model soon.

> **Try it:** Open Qwen Chat and upload a photo of a printed menu, sign or form in a language you don't read. Ask Qwen3.5 to translate it, then to explain anything unusual, such as a local dish or a legal term. Check one item against a dictionary to judge its accuracy.

## Can you run Qwen3.5 yourself, or should you use it online?

For most people, online is the answer. Even though only 17 billion parameters are active per token, all 397 billion must be loaded into memory. At 16-bit precision that is roughly 800 gigabytes, far beyond a laptop or a single graphics card. The model card gives instructions for serving it with tools such as vLLM, SGLang and Hugging Face Transformers, which are meant for multi-GPU servers.

The easy routes are Qwen Chat in a browser, or Alibaba Cloud Model Studio for developers who want an API. If you call it through an API, you control whether it thinks first:

```prompt
Summarize this 30-page report in ten bullet points for a busy
manager. Then list three questions the report leaves unanswered.
```

For a long report like this, leave thinking on. For quick tasks such as rewording a sentence, the model card shows how to turn thinking off for faster, cheaper replies. If you do have access to a large server, the open license means your organization can host Qwen3.5 itself and keep every document in-house.

## Sources

- Qwen3.5-397B-A17B model card: https://huggingface.co/Qwen/Qwen3.5-397B-A17B/blob/7cad2bae11cb/README.md
- Alibaba unveils new Qwen3.5 model for 'agentic AI era': https://www.timeslive.co.za/news/sci-tech/2026-02-16-alibaba-unveils-new-qwen35-model-for-agentic-ai-era/