TL;DR: Qwen is Alibaba's model family, and there are two ways to use it. Download the open-weight models and self-host them for free under Apache 2.0, or call the paid Alibaba Cloud API. Open models run from 0.6B to 480B parameters. The cheapest API tier, Qwen-Flash, is $0.05 input and $0.40 output per million tokens.
Qwen splits into two halves that most guides mix up. The first half is a set of open-weight models you can download and run on your own hardware, all under the permissive Apache 2.0 license. They range from a 0.6B model that fits on a phone to the 480B Qwen3-Coder. The second half is a paid API on Alibaba Cloud Model Studio, with three main tiers: Qwen-Max is the flagship, Qwen-Plus is the balanced default, and Qwen-Flash is the cheap high-volume option.
Qwen is the most downloaded open model family in the world, and its open weights let a team run a near-frontier model with zero per-token cost and full data control. But self-hosting needs real engineers. This guide helps your team pick between open weights and the API, the same way we match engineers to roles.
Key takeaways
- Two ways to run Qwen: free open weights you self-host, or the paid Alibaba Cloud API.
- Every open model is Apache 2.0. Sizes run from 0.6B on a phone to the 480B Qwen3-Coder.
- API tiers: Qwen-Flash ($0.05/$0.40) for volume, Qwen-Plus for the default, Qwen-Max for the hardest work.
- MoE models like Qwen3-235B-A22B use 22B active parameters, so they run far cheaper than their total size.
- Self-hosting means $0 per token but real infra cost. The trade is control and scale versus engineering effort.
The Qwen API tiers at a glance
Here are the main paid models on Alibaba Cloud Model Studio, with the model name your engineers will call. Use this table as your reference for the hosted API. The open-weight models come next. Prices are input then output, per one million tokens.
| Model | Context | Price in / out (per 1M) | Best for |
|---|---|---|---|
| Qwen-Max | 256K | $1.60 / $6.40 | Flagship, hardest reasoning and agents |
| Qwen-Plus | 1M | $0.40 / $1.20 | The balanced default for production |
| Qwen-Flash | 1M | $0.05 / $0.40 | High-volume simple tasks, cheapest |
| Qwen-Turbo | 131K | $0.05 / $0.20 | Fast and cheap, shorter context |
| Qwen3-VL-Plus | 262K | $0.20 / $1.60 | Vision: images, screens, documents |
| text-embedding-v4 | — | $0.07 / — | Embeddings for search and RAG |
How do you want to run Qwen?
Pick your setup to see the right model and the team you need.
For chatbots, tagging, and routing at scale, the API is fast to ship and Qwen-Flash starts at $0.05/$0.40 per million tokens. You pay no infra cost. See what building this in-house costs on our developer rate card.
For full data control and zero per-token cost, self-host a Qwen3 open model on your own GPUs. This needs strong deployment skills. Hire an AI model deployment engineer who runs vLLM and SGLang in production.
For agentic coding, Qwen3-Coder 480B matches top closed models and is open-weight, or you can call it via the API. Pair it with engineers who ship agents. Hire AI agent developers who build with it.
The model is the easy part. The hard part is the team that deploys and ships with it. We match you with pre-vetted engineers in about 24 hours. Tell us what you need.
The two ways to use Qwen
This is the choice that shapes everything else. Open weights mean you download the model and run it yourself. You pay for the GPUs, not for tokens, and your data never leaves your servers. The API means Alibaba runs the model and you pay per token. You ship faster and skip the infra, but you give up some control.
Most teams start on the API to test an idea, then move heavy or sensitive workloads to self-hosted open weights once the volume justifies the infra. A team doing millions of calls a day, or handling private data, often saves a lot by self-hosting. A team shipping a first version usually starts with Qwen-Plus on the API. Both paths use the same model family, so you can switch later.
The open-weight models
The open side is where Qwen stands apart. Every model below is Apache 2.0, free to download from Hugging Face or ModelScope, and free to run for any use including commercial. The lineup covers dense models and Mixture-of-Experts models. A dense model uses all its parameters on every token. An MoE model only activates a small slice, so it runs far cheaper than its total size suggests.
| Model | Parameters | Type | Good for |
|---|---|---|---|
| Qwen3-Coder 480B | 480B total / 35B active | MoE | The strongest open coding model |
| Qwen3-235B-A22B | 235B total / 22B active | MoE | Flagship open model, near frontier |
| Qwen3-30B-A3B | 30B total / 3B active | MoE | Strong, runs on a single GPU |
| Qwen3-32B / 14B / 8B | Dense | Dense | Fine-tuning and on-prem work |
| Qwen3-4B / 1.7B / 0.6B | Dense | Dense | Laptops, edge, and phones |
| Qwen3-VL | Multiple sizes | Vision | Open image and document reading |
The flagship open model, Qwen3-235B-A22B, scores close to the best closed models on hard reasoning and math benchmarks while using only 22B active parameters. Qwen3-Coder 480B matches top closed models on agentic coding tasks. Smaller dense models are the ones most teams fine-tune, because they are cheap to train and easy to deploy. Running any of these well is a job for strong AI and machine learning engineers.
The commercial API tiers
If you do not want to run servers, the Alibaba Cloud API gives you the same family, hosted. Qwen-Max is the flagship, tuned for the hardest reasoning and agent work. Qwen-Plus is the balanced default at $0.40 input and $1.20 output per million tokens, with a 1M token context window. Qwen-Flash is the cheap tier at $0.05 input and $0.40 output, built for high-volume simple tasks. Qwen-Turbo is a faster, shorter-context budget option.
The API also hosts the specialized models. Qwen3-VL-Plus reads images, screens, and documents. The text-embedding models power search and retrieval for a few cents per million tokens. There are also image and video generation models on the same platform. One API key reaches all of them, which is why some teams use the API even after they self-host their main text model.
Pricing compared across the lineup
On the API, output tokens are the real cost driver, and Qwen is cheap even at the top. Qwen-Turbo output is $0.20 per million. Qwen-Max output is $6.40 per million, the priciest hosted tier. That top rate still sits well below the flagship price of most Western labs. The chart below shows output price for the main API models.
Self-hosting changes the math completely. You pay $0 per token and instead pay for GPUs and the engineers who run them. For steady, high-volume work, that fixed cost often beats per-token pricing by a wide margin. For spiky or low-volume work, the API usually wins because you only pay when you call it. The right answer depends on your traffic, and getting it right is a real engineering decision.
Context windows and thinking mode
Qwen-Plus and Qwen-Flash both offer a context window around one million tokens, roughly 750,000 words. That is enough to drop a whole codebase or a long set of documents into one request. Qwen-Max uses a 256K window, and Qwen-Turbo a 131K window. The open models ship with 128K to 256K windows by default, and can be extended further with standard techniques.
Qwen3 models have a hybrid design with a thinking mode and a non-thinking mode. In thinking mode the model reasons step by step before it answers, which helps on hard math, code, and logic. In non-thinking mode it answers fast, which suits simple chat and tagging. You switch modes per request. Thinking mode uses more output tokens, so it costs more, and your engineers should turn it on only when a task needs it.
This matters for hiring. An engineer who understands thinking mode, MoE routing, and self-hosting will run Qwen far cheaper at the same quality. We see this gap often when we vet candidates. The ones who have shipped real open-model deployments know these levers cold. You can read more in our breakdown of Claude versus ChatGPT for coding.
Which model should you use
Start by choosing the path. If you need data control or run very high volume, self-host an open model. If you want to ship fast with no infra, use the API and default to Qwen-Plus. From there, move down to Flash for simple high-volume work and up to Max for the hardest tasks. The table below maps common jobs to a model.
| Job | Recommended model | Why |
|---|---|---|
| Support chatbot, FAQ | Qwen-Flash | Fast and cheap at high volume |
| Data tagging, classification | Qwen-Turbo or Qwen3-8B | Simple, repeated, low risk |
| Daily coding and code review | Qwen3-Coder | Top open coding model, API or self-host |
| Document summaries, extraction | Qwen-Plus | 1M context, good accuracy, low price |
| Private data, high volume | Qwen3-235B-A22B (self-host) | Full control, $0 per token |
| Hardest reasoning, agents | Qwen-Max | Flagship, worth the higher rate |
| Images and documents | Qwen3-VL | Strong vision, open or hosted |
A simple rule helps. Prototype on the API, measure your real volume, then decide if self-hosting pays off. We worked with a team that ran everything on a Western flagship API. Moving routine work to a self-hosted Qwen3 open model and keeping only the hardest calls on a premium API cut their model bill by more than half with no drop in quality.
Older Qwen models
Qwen ships fast, and older versions are still around. If your code calls an older model, it will usually keep working, but the newer versions are cheaper and stronger for the same job. The table shows the main older models and what to use instead.
| Older model | Status | Move to |
|---|---|---|
| Qwen2.5 open models | Active, superseded | Qwen3 open models |
| Qwen2.5-Max, Plus, Turbo | Active on API | Qwen-Max, Plus, Flash |
| Qwen2.5-Coder | Active, superseded | Qwen3-Coder |
| Qwen2-VL, Qwen-VL | Superseded | Qwen3-VL |
| Qwen1.5, Qwen 1.0 | Retired for new work | Qwen3 open models |
A move is usually small. On the API you change the model name. For open weights you download the new checkpoint and update your serving config. The Qwen model cards on Hugging Face and the Alibaba Cloud model list show every current and older version in one place, so check them before you migrate.
What this means for hiring
Knowing the lineup is step one. Getting value from it is the harder step. With Qwen the biggest lever is the deployment choice. The teams that win are the ones who know when to self-host, how to serve an MoE model efficiently, and how to route work between open weights and the API. That is engineering, not just a model choice.
Engineers who can deploy open models are scarce and costly in the United States. A senior AI engineer there runs $12,000 to $18,000 per month all in. The same skill from a pre-vetted engineer in Vietnam or the Philippines often costs $3,000 to $6,000 per month. The work is the same. The bill is not. Many of our clients build their whole open-model stack with an AI model deployment engineer from our network.
If you want to go deeper on the wider picture, our library of articles on AI tools and trends covers adoption, cost, and the open versus closed model debate. Read that next to see how fast this market is moving and why the right team matters more each quarter.
Compare other model families
Build your Qwen stack with the right team
Qwen gives you a rare mix: near-frontier open weights and a cheap hosted API. The model is the easy part. The team that deploys and ships with it is what creates value. We match you with pre-vetted senior engineers across Asia in about 24 hours, with no upfront cost and payroll handled.
