Skip to content
Neural Minute

Review · AI chatbots and companions

Qwen AI Review 2026: Alibaba’s Open Models Explained

By Maxime Aulne, Founder and Editor

Qwen chat is free and the API bills from $0.15 per million tokens. What Qwen4-Max does, the Apache 2.0 weights, and where the data question lands.

Minute Verdict

Qwen is the model family Alibaba Cloud builds and sells, and it arrives in two shapes. The chat app at chat.qwen.ai is free. The API, sold through Model Studio, is metered per token from $0.15 per million on qwen3.8-flash, while the flagship Qwen4-Max, released November 16, 2026, bills $1.5 per million input tokens and $4.5 per million output tokens.

That makes it the best value on the market for one kind of buyer: someone building a product who wants strong output, a 1M token context, and Apache 2.0 weights to fall back on if the price changes. A team that wants one seat price, no meter and no capacity planning is looking at the wrong product.

Best for: a developer shipping an agent or a document pipeline who can budget per token. Not for: a small team that wants a single monthly invoice per person. From: $0.15 per million input tokens, checked November 17, 2026

What Qwen does

Qwen is an assistant, a model catalogue and a weight release, all under one name. The assistant answers questions, searches the web and works on long documents. Behind it sits a family that spans text, vision, image generation, speech and video, sold through Alibaba Cloud Model Studio and, for most models, downloadable.

The current flagship is Qwen4-Max. It takes text, images and video as input and returns text, over a 1M token context with up to 131K tokens of output, or 262K when the model is reasoning. It handles function calling, structured JSON output, prefix completion and server side web search, and its endpoints speak the OpenAI and Anthropic formats, so a client written for either one can point at Qwen with a base URL and a key.

Below the flagship the range is wide enough to confuse. Qwen3.8-Max is the 2.4T parameter mixture of experts, Qwen3.7-Max the previous flagship, Qwen3.7-Plus the cost tier for document work, and Qwen3.8-Flash the model to pick when volume matters more than depth. Around them sit Qwen3-VL-Plus for screen and document reading, Qwen-Image-2.1, a 7B image model that generates and edits in one pass, and the Wan video models.

The part that separates Qwen from ChatGPT and Claude is the weights. Every open-weight release is Apache 2.0, from 0.6B to 235B-A22B, with documented routes through Ollama, llama.cpp, vLLM and SGLang, and support for more than 100 languages. A legal team that refuses a foreign API can run the same family on its own hardware, at a lower quality than Qwen4-Max.

On this page

What you get for the money

Model Input per 1M tokens Output per 1M tokens What the limit means
qwen3.8-flash $0.15 $0.47 1M token context, free quota of 1 million tokens
qwen3.7-plus $0.4 list $1.6 list 256K context at that rate, $1.2 and $4.8 above it
qwen4-max $1.5 $4.5 1M context, 131K output, video input
qwen3.8-max $2 $6 The 2.4T model, same window
qwen3.7-max $2.5 $7.5 The previous flagship, still the dearest text model
qwen-max $1.6 $6.4 Legacy model priced above the new flagship

The rate is per token, so the price of a job is the number of tokens it moves. A 30,000 token request costs about 0.45 cents on qwen3.8-flash at $0.15 per million, and 4.5 cents on qwen4-max at $1.5. That is the whole economics of the product: the flagship is ten times the price of the flash model for the same words.

Two discounts sit on top and both are dated as limited time. The plus tier carries 20 percent off its list price, and batch chat runs at 50 percent off on the original rates of $1.413 input and $4.24 output for qwen4-max. Batch file input is $0.707 and output $2.12, half of the interactive rate, which is the discount to take if the work can wait.

Caching is the other lever, and it only pays when a prefix repeats. Reading from the cache costs $0.075 per million tokens on qwen4-max, but writing to it costs $1.875 for a five minute cache and $3 for an hour. A fixed system prompt reused across thousands of calls pays for itself; a one off call does not.

Where it falls short

  • There is no seat price. Nothing on the model grid is sold as a flat monthly subscription, so a team that wants to know its bill in advance has to forecast tokens, or cap the account.
  • The free quota is geographically narrow. The 1 million free tokens per model, valid 90 days, are offered only on the Singapore deployment. Buyers routing traffic to the US, German or Japanese region start paying from the first call.
  • The old models cost more than the new ones. Qwen-max, a legacy entry, bills $1.6 input and $6.4 output, above the $1.5 and $4.5 of qwen4-max. Anyone who picks a model by name recognition overpays for a weaker model.
  • The plus tier steps up with context. Under 256K tokens of input the rate is $0.4 and $1.6; past that it becomes $1.2 and $4.8, so one long request moves the whole call into a rate three times higher.
  • Discounts are labelled limited time, and the 20 percent on the plus tier is attached to the price itself rather than to a dated promotion. Budgeting at the list rate is the safe reading.
  • Training on your prompts is opt-out, not opt-in. The training summary states that user interactions may be used to improve the service, and that an opt-out request has to be submitted.

Qwen alternatives

  • Claude, $20 a month on Pro billed monthly, the seat route for long documents and code, with no token meter to watch and no weights to run yourself: Claude review.
  • DeepSeek, from $0.003 per million input tokens on flash with a warm cache off-peak, the cheapest text meter in this catalogue, with no image or video input at that rate: DeepSeek review.
  • Kimi, $19 a month on Plus billed monthly, a seat based assistant instead of a meter, which suits one person who does not want to forecast tokens: Kimi review.

Who it is for: Qwen

Qwen fits a builder first. If you are shipping a product on top of an API, the combination of a 1M token window, video input, OpenAI compatible endpoints and a flash tier at $0.15 per million input tokens is hard to beat on cost per document. Add the Apache 2.0 weights and you keep an exit: the same prompts can move to a rented GPU if the rate or the data rules change.

It fits a second buyer too, the one who wants a free assistant and nothing else. The chat app asks for a login, takes images and files, and charges nothing, which puts it in the same bracket as the free tiers of ChatGPT and Claude.

It does not fit a company that needs a predictable invoice, a single per-seat price, or a contract with an American vendor. It also does not fit a team whose work depends on the flagship: at $4.5 per million output tokens, a long reasoning workload costs more than a seat on any competitor. The full list of chat tools sits in the AI chatbots category.

How we checked

This is a research review. We read the Model Studio price grid, the Qwen4-Max model card, the training data summary and the licence notice on the open weights on November 17, 2026, and we have not sent a request to the API or signed in to the chat app. No latency, no output quality and no censorship behaviour is verified here, and the reward for a hands-on pass would be exactly those three: time to first token on the flash tier, how the model handles a long document, and what it refuses. Our method page covers what a tested page adds, and our editorial policy covers how a page like this is funded.

What to do next

Open the free chat and put one real document in front of it before you consider the API, since the app costs nothing and answers the only question that matters at this stage. If the output holds, create a Model Studio account in Singapore to spend the 1 million free tokens on the flash tier, and move to qwen4-max only for the calls where quality is the bottleneck. File the training opt-out before the first confidential prompt, and set a spend cap on the account, because nothing in this catalogue stops you from running a $4.5 per million output model all night.

FAQ

Is Qwen free? The chat app at chat.qwen.ai is free to use. The API is metered, from $0.15 per million input tokens on qwen3.8-flash, and a new Model Studio account in Singapore gets 1 million free tokens per model for 90 days.

Who owns Qwen? Alibaba Cloud builds and runs it. The research is published by the Qwen team, the weights sit on Hugging Face and GitHub, and the commercial API is sold as Alibaba Cloud Model Studio.

Is Qwen open source? The open-weight models are licensed Apache 2.0, which allows commercial use and modification. The largest models, Qwen4-Max among them, are reachable through the API only, so open weights cover the family below the flagship.

How much does the Qwen API cost? On the International deployment, qwen4-max bills $1.5 per million input tokens and $4.5 per million output tokens, qwen3.8-max $2 and $6, qwen3.7-max $2.5 and $7.5, and qwen3.8-flash $0.15 and $0.47.

Is Qwen safe to use with client data? The API can be deployed in six regions, US (Virginia) included, which settles where the inference runs. The training summary says user interactions may be used to improve the service unless an opt-out request is filed, so a team with confidential material should file it before the first prompt.

Sources

Sources

Update log

Published November 17, 2026. Prices, quotas, context limits and licence read on November 17, 2026.

Newsletter

One email when a review lands.

That is when the promo codes are still fresh, before the prices move.

Nothing else. Unsubscribe in one click.