Qwen AI: free Qwen Chat or Qwen locally ?

#Qwen AI in brief: a service and a family of models

Qwen is the name of the family of large language models developed by Alibaba. That name covers three distinct things you need to tell apart before choosing anything.

Qwen Chat (chat.qwen.ai)

A web-based chat app, similar in principle to ChatGPT or Le Chat. You type in your browser, and the responses are computed on Alibaba's servers.

Alibaba Cloud's Qwen API

The same kind of models, but called programmatically and billed by usage. This is the channel for the largest models, such as Qwen 3.8-Max, available via API since August 3, 2026, whose weights have not been released.

Open-weight Qwen models

Model files published on huggingface.co/Qwen, which you download once and then run on your own machine with Ollama, LM Studio, or llama.cpp. Recent example: Qwen3.8-27B, released on August 14, 2026 under the Apache 2.0 license.

So the question “Is Qwen AI free?” has two answers. Online, basic access to Qwen Chat doesn't require payment, but your messages go to Alibaba and usage remains subject to the service's terms. Locally, the model costs nothing and your data stays on your machine, but your hardware does the work, and only some variants are available for download.

#Qwen Chat: Alibaba's online assistant

Qwen Chat is the consumer interface for Qwen AI. You access it at chat.qwen.ai from any browser, with nothing to install. A selector lets you choose among several Qwen models; the list changes with each release and may include models that exist only as hosted versions.

The appeal for getting started with Qwen is obvious: no graphics card required, models far larger than a desktop PC can load, and a ready-made interface with conversation history. Qwen understands and writes French well, making it an accessible entry point for anyone searching for "Qwen AI in French."

The trade-off is the same as with any online service. Every message and every attached file is sent to and processed on remote servers operated by Alibaba. As QuelLLM's overview of Chinese LLMs points out, for cloud services, the provider's jurisdiction and data retention policy really matter. Read the terms of use and privacy policy shown by the service before you paste in a contract, customer data, or an internal document.

Access

Browser, no installation. You may be asked to sign in to an account to keep your history or access certain features: check what the page shows when you open it.

Models

Whichever ones Alibaba chooses to offer. A model shown in Qwen Chat isn't necessarily released as open weights.

Data

Processed at Alibaba. Offline, nothing works.

Stability

The service can change its default model, features, or limits overnight, and you have no way to pin a version.

#Free Qwen AI: what's free and what isn't

The word "free" covers three very different situations depending on the Qwen channel you use.

1. 01
   Qwen Chat: free to use, subject to the service's terms

   You don't pay to chat, but access depends on usage rules set by Alibaba (volume limits, prioritization during periods of high demand, features restricted to signed-in accounts). These rules aren't a stable commitment. If a specific quota matters for your use case, check it yourself on the service's page and date your note: a figure read on a forum or in an old article is worthless.
2. 02
   Alibaba Cloud API: pay-as-you-go

   As soon as you want to call Qwen from a program, a script, or an automation hosted by Alibaba, you go through the Alibaba Cloud API, which is billed based on usage. This guide doesn't cover that pricing: check it directly with Alibaba Cloud, since it varies by model and region.
3. 03
   Local Qwen: free, no quota, but on your own hardware

   Open-weight models are free to download. Most recent variants are released under Apache 2.0, which also allows commercial use; still, check the LICENSE file in the specific repository. There's no account, no message limit, and no cutoff when the service is overloaded. The real cost is your machine's memory, electricity, and setup time.

The local Ollama API, which listens on http://localhost:11434, also gives you the equivalent of a free Qwen API for your own scripts: as long as the machine is running, you can call it as much as you want, within the limits of what your hardware can compute.

#Hosted model or open weights: the distinction that saves you from bogus tutorials

This is the biggest source of confusion around Qwen AI. Alibaba regularly releases a flagship version, often called "Max," first or exclusively via the API and in Qwen Chat, then publishes smaller open-weight variants. The generation name is the same; the capabilities are not.

Qwen 3.8 is the most recent example. Qwen 3.8-Max, presented as a MoE with roughly 2,400 billion parameters and a context of one million tokens, has been available via API since August 3, 2026; its weights are not distributed, and it wouldn't fit on any consumer machine anyway. Qwen3.8-27B, a dense 27-billion-parameter model, was released on Hugging Face on August 14, 2026 under Apache 2.0, with an official Ollama tag the same day. Both carry the "Qwen 3.8" name, but only one can be installed on your own machine.

Hosted only

Available through Qwen Chat or the API. No files to download. Any page claiming to install it "locally" is mistaken or is talking about a different model.

Open weights

Public repository at huggingface.co/Qwen, model card, license file, and often a Ollama tag or GGUFs ready for LM Studio.

Announced, not yet released

Mentioned at a conference or in a blog post, with no repository. Between August 3 and 13, 2026, Qwen3.8-27B was in this state, even though many articles were already presenting it as installable.

#Checking for yourself that a Qwen variant downloads

The test is simple: can you download the model right now, with a single command, without a waitlist? Three checks are enough, and they remain valid for future generations, including against this page once it's out of date.

1. 01
   Find the repository on Hugging Face

   Go to huggingface.co/Qwen and search for the exact variant name (for example, Qwen3.8-27B). A public repository with weight files and a model card is proof of release. A collection or announcement page alone isn't enough.
2. 02
   Read this repository's license

   The LICENSE file or the model card header states the terms. Apache 2.0 is common at Qwen, but some older or specialized variants may carry a custom license: the specific repository is what counts, not the family's reputation.
3. 03
   Test the tag in Ollama

   Find the model on ollama.com/library, then start the download. If the tag doesn't exist, Ollama returns an error instead of downloading anything. Once the model is pulled, ollama show affiche its specs (parameters, quantization, context, license).

# 1. Lister les dépôts publics de l'organisation Qwen contenant « 3.8 »
curl -s "https://huggingface.co/api/models?author=Qwen&search=3.8" | grep -o '"id":"[^"]*"'

# 2. Télécharger le modèle si le tag existe (erreur sinon)
ollama pull qwen3.8:27b

# 3. Afficher ses caractéristiques une fois téléchargé
ollama show qwen3.8:27b

If the first command returns no repository for an announced variant, it hasn't been released: there's no point looking for an installation tutorial. If it returns a repository but no Ollama tag exists yet, community GGUF files often appear on Hugging Face and can be loaded into LM Studio; in that case, check who produced them before using them.

#Which open Qwen model fits your machine?

Locally, the deciding factor isn't the model's generation but its size compared with your graphics card's memory (VRAM) or your Mac's unified memory. With Q4_K_M quantization, the best quality/size trade-off for getting started, QuelLLM's benchmarks are as follows: 3B ≈ 2 GB, 7B ≈ 5 GB, 14B ≈ 9 GB, 32B ≈ 19 GB, 70B ≈ 40 GB. Add 1 to 2 GB for context and the system.

8 GB of VRAM or a 16 GB Mac

A Qwen with 4 to 9 billion parameters, for example qwen3.5:4b or qwen3.5:9b in Ollama. Good enough for chat, rephrasing and simple questions in French.

12 GB (RTX 3060 12 GB, RTX 4070)

A model around 14B in Q4_K_M, or a 9B in Q8_0 with headroom for a longer context.

16 GB (RTX 4080)

The upper end of mid-size models, or a MoE that partially spills over into system memory, at lower throughput.

24 GB (RTX 4090) or a Mac with 32 GB or more

Qwen3.8-27B, which weighs 18 GB at Q4_K_M according to the Ollama library, or the Qwen3.6 35B-A3B MoE. This is the tier where local comes closest to an online assistant for everyday use.

Need to analyze images

Qwen3-VL, the vision variant, for example qwen3-vl:8b. A dedicated guide covers how to install it and its limitations.

Don't try to reproduce the largest Qwen Chat model locally: it's neither possible nor necessary. A 9B to 27B model is enough for a good share of everyday writing, summarization, and coding tasks; judge for yourself using your own examples. For tasks that truly require the flagship model, the online service remains the answer, provided your data can go there.

#Install Qwen locally with Ollama or LM Studio

Two free tools cover the essentials. Ollama runs in the background and is controlled from the command line or via API; it's often paired with Open WebUI to get an interface similar to Qwen Chat. LM Studio is an all-in-one desktop app with built-in model search and a chat window. Here's the shortest path.

1. 01
   Install the tool

   Download Ollama from ollama.com or LM Studio from lmstudio.ai, for Windows, macOS, or Linux. Ollama starts a local service that listens on http://localhost:11434.
2. 02
   Choosing the size based on your memory

   Go back to the guidelines in the previous section. If in doubt, start with a 4B to 9B variant: it downloads quickly and lets you judge its quality in French.
3. 03
   Download and run the model

   With Ollama, a single command fetches the weights and then opens the chat in the terminal. In LM Studio, type “qwen” in the model search, pick a Q4_K_M GGUF version of the right size, then load it in the chat tab.
4. 04
   Verify that everything is local

   Once the model is downloaded, disconnect from the network and keep chatting: responses still come through. That's the concrete difference from Qwen Chat.
5. 05
   Connect your tools (optional)

   The Ollama API lets you use Qwen from a script, a code editor or Open WebUI, with no key and no quota.

# Petite machine (8 Go de VRAM ou Mac 16 Go)
ollama run qwen3.5:9b

# Carte 24 Go ou Mac 32 Go et plus : Qwen3.8 27B (18 Go en Q4_K_M)
ollama run qwen3.8:27b

# Lister les modèles installés
ollama list

# Appeler Qwen depuis un script via l'API locale
curl http://localhost:11434/api/generate -d '{
  "model": "qwen3.5:9b",
  "prompt": "Explique en trois phrases ce qu’est un modèle à poids ouverts.",
  "stream": false
}'

Tag names change quickly: before downloading, check ollama.com/library for the exact variant name and its size in GB. Installing Qwen3.8-27B in detail (quantizations, context, reasoning mode) is covered in a separate guide; this page sticks to choosing between online and local.

#Qwen Chat or local Qwen: how to decide

There's no universally right choice. The two options complement each other, and many users keep both: the online service for one-off, low-stakes requests, and the local model for anything involving sensitive data or automation.

Choose Qwen Chat if…

you want to try Qwen AI in two minutes, you don't have a graphics card, you need the largest model in the family, and your messages contain nothing confidential.

Choose Qwen locally if…

your documents must not leave your machine, you want a pinned model whose version doesn't change without your say, you work offline, or you automate tasks and don't want to depend on any quota.

The limits of running locally

Quality depends on the model size your machine can load; hosted flagship models remain out of reach; installation and updates are up to you.

Limits of the online service

Data sent to a third party, terms and usage limits that can change, and dependence on the service's availability and your connection.

A simple way to decide: ask Qwen Chat and a local Qwen of the size your machine can handle the same question, using an example representative of your real use but without sensitive data. If the local answer is good enough, you gain privacy and no quotas. If the gap is too wide, you know exactly what you're trading off by staying online.

#Go further

This page helps you choose between the online assistant and running locally. To get started or to compare Qwen with other model families, these guides pick up from here.