Which is better, Qwen or DeepSeek?
Artificial Analysis runs each model through the same ten tests, from Terminal-Bench 4.0 to Humanity's Last Exam, and works out what one task costs at list price. That cost counts the tokens a model spends while it thinks. Alibaba's top model, Qwen3.8 Max, scores 45. DeepSeek's best, V4.1 Flash, scores 39. So you pay 20 times as much per task for the six extra points.
At the cheap end of each family, it's a dead heat. Qwen3.8 Flash Next scores 40 for $0.37 a task, and DeepSeek V4.1 Flash scores 39 for $0.27. What sets them apart there is the clock. Artificial Analysis times every task as well, and it took DeepSeek's Flash 289 seconds on average to finish one. It took Qwen's Flash Next 1,809. Qwen3.8 Max needed 2,508, which is about 42 minutes.
| Model, read 12 October 2026 | Intelligence Index | Cost per index task | Input / output per 1M tokens | Output speed | End to end response | Weights |
|---|---|---|---|---|---|---|
| Qwen3.8 Max (0902) | 45 | $5.41 | $2.00 / $6.00 | 36 tokens/s | 71.9 s | Closed |
| Qwen3.8 Flash Next | 40 | $0.37 | $0.15 / $0.47 | 51 tokens/s | 51.3 s | Qwen Community License 1.0 |
| Qwen3.8 2.4T A95B | 40 | $2.16 | $2.00 / $6.00 | 37 tokens/s | 71.4 s | Qwen3.8-Max License |
| DeepSeek V4.1 Flash (max) | 39 | $0.27 | $0.30 / $1.20 | 211 tokens/s | 12.8 s | MIT |
| DeepSeek V4 Pro 0813 (max) | 36 | $0.67 | $1.32 / $3.96 | 93 tokens/s | 28.5 s | MIT |
| Qwen3.8 27B (xhigh) | 34 | $1.01 | $0.50 / $3.00 | 47 tokens/s | 57.5 s | Apache 2.0 |
| Qwen3.7 Plus | 25 | $0.22 | $0.40 / $1.60 | 55 tokens/s | 48.0 s | Closed |
All of the scores, costs, speeds and response times come from Artificial Analysis (Intelligence Index v4.3.2), and we read them on 12 October 2026. The prices are what each maker charges on its own API: Alibaba's Model Studio for Qwen, and DeepSeek's API for DeepSeek. The license column is from each model's card on Hugging Face. Per token, Qwen's Flash Next costs about half of what DeepSeek's Flash does. Per task it costs more. Artificial Analysis counts every token a run uses, cached input included, and DeepSeek bills a cached input token at just $0.006 per million at peak.
Is Qwen or DeepSeek better for coding?
Qwen3.8 Max is, if your agent works in a terminal. On Terminal-Bench 4.0, which is one of the ten tests in the index, it scores 39% and DeepSeek V4.1 Flash scores 27%. On SciCode, which asks for scientific functions from a written spec, the two of them tie at 52%.
If you drop to the cheap models, DeepSeek is a little ahead. Its Flash scores 27% on Terminal-Bench to the 25% of Qwen3.8 Flash Next, and it's 52% to 51% on SciCode. DeepSeek's Flash also leads on AutomationBench-AA with 69%, where both of the Qwen models score 56%. The two big open models are close to each other. DeepSeek V4 Pro scores 14% on Terminal-Bench, and Qwen3.8 2.4T A95B scores 11%.
Speed is the part that a coding agent feels the most. DeepSeek's API wrote 211 tokens a second for V4.1 Flash, and Alibaba's wrote 36 for Qwen3.8 Max. On a job that loops for an hour, that's the difference between waiting on it and not. Claude vs DeepSeek runs the same tests against a closed model.
Which open weights can you build a business on, Qwen or DeepSeek?
DeepSeek's, and there's nothing for you to sign. Every DeepSeek V4 release on Hugging Face, from V4 Flash and V4 Pro in April 2026 to V4.1 Flash in September, comes under the plain MIT license. Qwen's three August 2026 language models come with three different licenses, and we read each of the LICENSE files on 12 October 2026.
| Open model, as of October 2026 | License | Who needs a separate license | When the model name must show in your app |
|---|---|---|---|
| DeepSeek V4.1 Flash | MIT | Nobody | Never |
| DeepSeek V4 Pro 0813 | MIT | Nobody | Never |
| Qwen3.8 27B | Apache 2.0 | Nobody | Never |
| Qwen3.8 Flash Next | Qwen Community License 1.0 | Any company running a Model as a Service or AI Work Assistant business, at any revenue | Above 100M monthly users or $20M monthly revenue |
| Qwen3.8 2.4T A95B | Qwen3.8-Max License | The same businesses, once group revenue passes $50M in 12 months | Above 100M monthly users or $20M monthly revenue |
| Qwen3.8 Max (0902) | No weights released | Not applicable | Not applicable |
The Qwen Community License says a Model as a Service is when you give a third party inference or fine tuning through an API or a hosted endpoint. An AI Work Assistant is a product that's built mainly for AI coding or office work, and the license names Qoder and QwenWork as examples. So if a startup sells a coding tool that runs on Qwen3.8 Flash Next, it needs a deal with Qwen before it can use the model for anything commercial. Internal use is exempt, as long as no one outside the company gets the model or what it writes.
The same startup on DeepSeek's weights only has to keep the MIT notice in its copies. If all you'll do is chat with the model, none of this applies to you. The open source LLMs guide covers what it takes to run weights like these.
How much do the Qwen and DeepSeek APIs cost in October 2026?
DeepSeek V4.1 Flash costs $0.30 for each million tokens you send and $1.20 for each million that it writes. For V4 Pro it's $1.32 and $3.96. DeepSeek's pricing page cuts both of them in half off peak, and its peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays. On Chinese public holidays it's off peak all day.
Alibaba's Model Studio price list, which it updated on 11 October 2026, puts qwen3.8-max at $2 and $6 in its Singapore region, and qwen3.8-flash at $0.15 and $0.47. Qwen3.7 Plus is $0.40 and $1.60 for a prompt of up to 256K tokens, and it costs more above that. If you open a new account in Singapore, you get 1M free tokens for each model, and they last 90 days. DeepSeek's page doesn't list a free quota.
Alibaba also sets its prices by region, which DeepSeek doesn't do. Its China (Beijing) prices are listed apart, and Qwen3.8 Max Prime is only sold there. So of the two Flash models, the one from Qwen is cheaper per token, and the one from DeepSeek is cheaper for each task it finishes.
Where do Qwen and DeepSeek keep your chats?
Qwen Chat keeps them in Singapore, and DeepSeek keeps them in China. The privacy policy for chat.qwen.ai names NTH Power Global Tech Singapore as the company that runs it, and it says: "We collect, process, and store your personal data in Singapore." Alibaba Cloud's company in Singapore runs the model for each of your requests. The same policy says Alibaba Cloud deletes your input once the answer is done.
DeepSeek's privacy policy says that it stores your personal data in the People's Republic of China. It uses that data to train its models, and it gives you a right to opt out of that.
| As of October 2026 | Qwen Chat (chat.qwen.ai) | DeepSeek app |
|---|---|---|
| Where your data is stored | Singapore | People's Republic of China |
| Chats used for training | De-identified content and feedback only, not raw input directly | Yes |
| Opt out of training | Not described in the policy | A listed right to opt out |
| Who runs the model | Alibaba Cloud (Singapore), which deletes input after each answer | DeepSeek |
If you run the open weights of either model on a host you trust, your data goes to neither country. That's the one way to settle it for both of them.
Should you pick Qwen or DeepSeek?
Pick Qwen when the top score or where your data sits decides it. Pick DeepSeek when cost per task or speed does.
| Pick Qwen if | Pick DeepSeek if |
|---|---|
| You want the top score in either family (45 for Qwen3.8 Max) | Cost per task decides it ($0.27 for V4.1 Flash against $5.41) |
| Your terminal agent needs the best Terminal-Bench 4.0 result (39% against 27%) | You want fast answers (211 tokens a second against 36) |
| You need video input on a cheap model (Qwen3.8 Flash Next reads text, images and video) | You'll sell an API or coding product on the weights (MIT, no separate license) |
| Your data has to stay out of China (Qwen Chat stores it in Singapore) | You need a 1M window on the cheap model (1M against 256K) |
| You want a 27B model under Apache 2.0 | Your jobs can wait for off peak rates at half price |
Whizi has both of these families, but its Pro plan keeps the best of Qwen back. Pro is $29.99 a month with 2,500 credits, and it has DeepSeek V4.1 Flash and Qwen3.7 Plus at 1 credit a message each. Its side by side view sends one prompt to both of them. But Qwen3.7 Plus scores 25, so on Pro the DeepSeek model is the one to pick. Qwen3.8 Max (6 credits), Qwen3.8 2.4T A95B (8), Qwen3.8 Flash (1) and DeepSeek V4 Pro 0813 (2) are on Powerhouse at $49.99. Starter includes neither family. The Qwen and DeepSeek docs list every row.
- Use DeepSeek V4.1 Flash when cost per task or speed decides it
- Use Qwen3.8 Max when the top score matters more than a 42 minute task
- Read the Qwen license before you sell an API or coding product on Qwen3.8 Flash Next
- Use Qwen Chat over the DeepSeek app if your data has to stay out of China
- Batch DeepSeek API work into off peak hours for half price
Frequently asked questions
Is Qwen better than DeepSeek?
On score, yes. On cost and speed, no. Read 12 October 2026, Qwen3.8 Max scores 45 on the Artificial Analysis Intelligence Index to the 39 of DeepSeek V4.1 Flash. But it costs $5.41 a task to DeepSeek's $0.27, and it writes 36 tokens a second to DeepSeek's 211. At the cheap end of each family, Qwen3.8 Flash Next and DeepSeek V4.1 Flash are just one point apart. Kimi vs DeepSeek runs the same score and cost comparison against Kimi.
Is Qwen or DeepSeek better for coding?
Qwen3.8 Max is better for terminal agents, with 39% on Terminal-Bench 4.0 to the 27% of DeepSeek V4.1 Flash. They tie at 52% on SciCode. If you'll sell a coding product built on open weights, DeepSeek's MIT license needs no extra deal, and the license on Qwen3.8 Flash Next does. For a wider coding comparison, see coding picks beyond ChatGPT.
Is Qwen cheaper than DeepSeek?
Per token, yes. Alibaba lists qwen3.8-flash at $0.15 in and $0.47 out per million tokens in Singapore, and DeepSeek lists V4.1 Flash at $0.30 and $1.20. Per finished task, DeepSeek is cheaper, at $0.27 to $0.37 on Artificial Analysis. DeepSeek also halves its rates off peak. LLM API pricing statistics gives median prices across paid text models.
Can I use Qwen and DeepSeek in one app?
Yes. Whizi Pro at $29.99 a month has DeepSeek V4.1 Flash and Qwen3.7 Plus at 1 credit a message each, with a side by side view. Qwen3.8 Max is on Powerhouse at 6 credits a message. Starter includes neither family. One subscription for all AI models compares other apps that bundle models.
