What is Qwen, and why is Alibaba making AI models
Five Qwen models sit in the MultiChats picker. Here is what they handle well, what they quietly refuse, and the moment to switch to something else.
Start from the job in front of you
Open the model list in MultiChats and five current entries appear under Alibaba, every one of them called Qwen and followed by numbers that need explaining: Qwen3 235B, Qwen3 VL 235B, Qwen3.5 397B, Qwen3.6 35B and Qwen3.8 27B. Qwen3.5 9B still shows with a retirement notice until 21 October 2026. Learning that Qwen is a family of models from a Chinese technology company tells you nothing useful about whether to tap one of them.
So bring a real job to it, and keep the same job for the rest of this article. Say you are renting a flat, and the whole arrangement has been negotiated over six weeks of email. Forty or so messages, three people, an agent who changes the deposit figure twice and a landlord who answers questions in the wrong order. You want three things out of that pile: what was agreed, what is still open, and what you owe and when. The thread is long, it is entirely text, and every fact you need is already sitting inside it. That is a task with a shape, and the shape is what decides which model deserves it.
Qwen belongs to a cloud company
The reason there are five of them, and the reason a new one seems to land every few weeks, is that the models are a means to something else. Alibaba describes itself this way on its own corporate site:
Now, Alibaba Group is a global technology company focused on AI + Cloud and commerce. We empower consumers and enterprises with our full-stack AI capabilities and services, from applications to compute infrastructure. Our AI technology based on the Qwen family of large language and multimodal models powers the intelligence behind our services across enterprise solutions and consumer platforms.
That paragraph answers the second half of the question in the title. Alibaba sells cloud computing and it sells commerce, and the models are what make both of those worth buying. On Alibaba Cloud's own platform, Model Studio, the Qwen line is sold as a paid API in three tiers called Max, Plus and Flash. Inside the consumer business, the same family runs the shopping assistant that was connected to Taobao's entire product catalogue in May 2026, and Alibaba's milestone page says its consumer-facing Qwen passed 300 million monthly active users in February 2026. Meanwhile the Qwen team, described on Hugging Face as "the large language model family built by Alibaba Cloud", has published in the region of 465 model checkpoints on that site since the first Qwen-7B release in August 2023, a good many of them under the Apache 2.0 licence.
A research lab publishes to prove a point. A cloud company publishes to be picked up, deployed, benchmarked and eventually run on infrastructure it would like to rent you. That is why the cadence is relentless, and it is also why any model picker, ours included, holds a snapshot of that family rather than the whole of it. Alibaba's newest language models carry the Qwen3.8 name, and since 4 October 2026 one of them, Qwen3.8 27B, has sat in our picker next to entries from the 3, 3.5 and 3.6 generations.
What the free Qwen does with the thread
Two of the five sit on the free plan: Qwen3 235B and Qwen3 VL 235B. For a forty-message email thread, the first one is the interesting entry.
Its published model card advertises 262,144 tokens of context natively, which is a great deal of email, so the negotiation goes in whole rather than in slices. Paste it into the message box and it is there. (Attaching it as a text file is a paid-plan feature, so on the free plan pasting is the route.) Quality does taper before any model's advertised maximum, which is why the context meter in a long thread starts telling you where that model's answers tend to get shaky, and it is worth believing.
What the model is good at here is following an instruction exactly as given. "List every figure anyone named for the deposit, with who said it and the date" comes back as that list, without a detour into advice you did not ask for. You will also notice it does not think out loud. Its card states plainly that it supports only non-thinking mode and produces no reasoning blocks, and our entry matches, so no reasoning panel appears above the answer. The three Qwen entries on the Pro plan behave the other way round: reasoning is switched on for all of them and cannot be turned off, so even a one-line question gets a visible round of deliberation first.
One caution belongs here rather than at the end, because it changes how you read the reply. With no reasoning pass on screen there is nothing between your question and the answer, and so no visible place where a wrong turn would show itself before a figure arrives. On a negotiation where the deposit moved twice, that is the difference between an answer and a claim. Ask for the quoted sentence and the date alongside every number, then check two of them against the thread yourself. Any model that hands you a figure without showing where it came from should be made to produce the receipt.
When the thread stops being text
Three things tend to happen next in a rental negotiation, and each one is a reason to leave Qwen for that message.
The agent sends a screenshot of the payment schedule. Qwen3 235B reads no images at all, and the app says so out loud: "This model can't read images. Switch to a model with vision to send images." The other free entry, Qwen3 VL 235B, does read them, and its own card claims text recognition across 32 languages in poor light and at an angle. Worth knowing that a card describes the model as published, while a picker describes what an app has wired up: we accept images on that entry, and not video.
The landlord sends the contract as a PDF. No Qwen entry in our picker accepts a PDF file, and the attachment is refused with much the same message. 24 of our 59 models do accept one, so that document belongs in a conversation with one of them.
Then you want to check whether the deposit protection rules the agent quoted are still current. None of the five Qwen models can search the web, and this is the one worth understanding properly, because the search button is not always absent. Set a preferred web search provider in Settings, under Tool Preferences, switch Intent Detection off in that same place, and the toggle appears on every model, these included. What follows is a swap: the message gets answered by that search model instead, so the reply on your screen came from somewhere other than Qwen. Useful behaviour, as long as you know you asked for it. With Intent Detection left on, as every new account has it, a Qwen message never reaches search at all.
None of the five run voice mode either, so talking your way through the thread means a different model again.
Is Qwen safe to use?
That phrase is one of the most common searches attached to this name, and it deserves an answer rather than reassurance.
Requests to Qwen models from MultiChats never reach Alibaba. Every open-weight model in this part of our catalogue carries a routing restriction written into the registry itself, which confines the request to a vetted set of US and EU hosts and shuts out the model creator's own endpoints. It is enforced where it cannot be forgotten: a test fails our build if such a model is ever added without the flag.
On training, the accurate version has three parts. We do not train on your conversations. Some providers might, depending on the model you choose. Where we know of a difference, the model's info card carries a notice in that provider's own words, which is how one card can read "Routed via OpenRouter to third-party providers that may briefly retain requests; they do not train on your data" while Grok's says outright that "xAI may use your conversations to train Grok models." The Qwen entries carry no such note, because the routing restriction is what removes the caveat.
Our pick, and when we would drop it
For the email thread as described, we would run it on Qwen3 235B and ask for every number with its quoted sentence attached. It holds the whole negotiation without slicing it up, and it follows a precise instruction precisely, with no preamble in front of the answer. That is a good fit, and fit is the only thing worth choosing on.
We would drop it the moment the job changes character. A screenshot sends you to the vision entry and a contract PDF sends you out of the family altogether, while anything that turns on what is true today rather than what is in the thread wants a model that can go and look. The five Qwen entries are strong at reading closely what you hand them and unremarkable at everything that requires reaching outside it, and working out which of those two jobs you actually have helps narrow your choice. Three of the five sit behind Pro and the two that suit this particular job do not, which means the only thing standing between you and disagreeing with all of the above is one long, annoying thread.