Qwen vs DeepSeek: two open-weight model families for agent builders
Qwen3.8-Max costs $2 / $6 per 1M input / output tokens on Alibaba Cloud Model Studio and DeepSeek V4 Pro costs $1.32 / $3.96 at peak and half that off-peak, both accept 1M-token inputs, and both families publish open weights under different licences, so the choice turns on licence terms, hosting region and results on the team's own eval.
Checked
Qwen and DeepSeek are the two open-weight model families most often compared by teams that want lower token prices or the option to run a model themselves. Both sell hosted APIs and both publish weights on Hugging Face.
They differ in how the hosted API is sold and where it runs, and in what the open licence allows. Those differences matter more to a team building an agent than headline benchmark scores, which come from each vendor's own tables and cannot be compared across them.
On the API: price and context
On Alibaba Cloud Model Studio, qwen3.8-max costs $2 / $6 per 1M input / output tokens, qwen3.7-plus $0.40 / $1.60 on requests up to 256K input tokens and $1.20 / $4.80 above that, and qwen3.8-flash $0.15 / $0.47. These models accept up to 1M input tokens per request.
DeepSeek V4 Pro costs $1.32 / $3.96 at peak and $0.66 / $1.98 off-peak, and DeepSeek V4.1 Flash $0.30 / $1.20 at peak and $0.15 / $0.60 off-peak. Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays. Both DeepSeek models take 1M tokens of context with up to 384K output, and accept both the OpenAI and Anthropic request formats.
Open weights and licences
Qwen
Qwen3.8-27B is published under Apache 2.0. The large open model, Qwen3.8-2.4T-A95B (2.4T total, 95B active), is published under a model-specific licence tagged qwen3.8-max, which should be read before commercial use. Its open context is 262,144 tokens natively, extensible to about 1M. Alibaba describes the hosted Qwen3.8-Max as the official version of that model with more features, including vision input and built-in tools.
DeepSeek
DeepSeek V4 Pro's weights are published under the MIT licence: 1.6T total and 49B active parameters, with a 1M-token context.
Self-hosting
Both flagship open models need a multi-GPU cluster to serve. The 27B Qwen model is the smaller option for teams that want to self-host on modest hardware and can accept a smaller model.
Where hosted data is processed
Alibaba Cloud Model Studio lists deployment regions including Singapore, Virginia, Frankfurt, Tokyo, Hong Kong and Beijing, and charges the same qwen3.8-max price in Singapore and Virginia. A team can pick a region that matches its data rules.
DeepSeek's privacy policy states that it collects, processes and stores personal data in the People's Republic of China. Teams that need another location can self-host DeepSeek's MIT-licensed weights or use a host in their own region.
How to decide
If the team wants a hosted API in a specific non-Chinese region, Qwen on Model Studio offers that directly. If it wants the lowest flagship token price and can schedule work off-peak, DeepSeek's hosted API is worth testing. If it will self-host, the licence decides as much as the model: MIT for DeepSeek, Apache 2.0 for Qwen3.8-27B, and a model-specific licence for the large Qwen weights.
Motiontic's engineers are vendor-neutral. They compare candidate models on the client's own data with an eval suite, route each step of the agent to the cheapest model that passes its eval, and leave the eval suite with the client, so the next model change is a measured decision rather than a guess.
Qwen and DeepSeek for teams building agents. Prices in USD per 1M input / output tokens, as of 1 November 2026
| Qwen (Alibaba) | DeepSeek | |
|---|---|---|
| Flagship hosted model | qwen3.8-max, $2 / $6 | DeepSeek V4 Pro, $1.32 / $3.96 peak, $0.66 / $1.98 off-peak |
| Small hosted model | qwen3.8-flash, $0.15 / $0.47 | DeepSeek V4.1 Flash, $0.30 / $1.20 peak, $0.15 / $0.60 off-peak |
| Input limit / context | Up to 1M input tokens per request | 1M context, up to 384K output |
| Large open model | Qwen3.8-2.4T-A95B: 2.4T total, 95B active | DeepSeek V4 Pro: 1.6T total, 49B active |
| Licence of large open model | Model-specific (tagged qwen3.8-max) | MIT |
| Hosted regions | Singapore, Virginia, Frankfurt, Tokyo, Hong Kong, Beijing and others | Privacy policy: data stored in the People's Republic of China |
| Request formats | Model Studio API | OpenAI and Anthropic formats |
Common questions
Is Qwen or DeepSeek cheaper?
At the flagship tier, DeepSeek V4 Pro at $1.32 / $3.96 per 1M input / output tokens at peak (half that off-peak) costs less than qwen3.8-max at $2 / $6. At the small end, qwen3.8-flash at $0.15 / $0.47 costs less than DeepSeek V4.1 Flash at peak, and matches its off-peak input price with a lower output price.
Are Qwen and DeepSeek open source?
Both publish open weights. DeepSeek V4 Pro is under the MIT licence. Qwen3.8-27B is under Apache 2.0, while the large Qwen3.8-2.4T-A95B model uses a model-specific licence that should be read before commercial use.
Can Qwen be hosted outside China?
Yes. Alibaba Cloud Model Studio lists regions including Singapore, Virginia, Frankfurt and Tokyo, and qwen3.8-max is priced the same in Singapore and Virginia. Both Qwen and DeepSeek weights can also be self-hosted.
Is the hosted Qwen3.8-Max the same as the open weights?
Not exactly. Alibaba describes Qwen3.8-Max as the official version based on Qwen3.8-2.4T-A95B with more features, such as vision input, a non-thinking mode, a 1M default context and built-in tools. The open model is text-only and runs in thinking mode.
How should a team choose between Qwen and DeepSeek?
By fixing the hosting and licence constraints first, then scoring the remaining options on the agent's real tasks with an eval suite. Motiontic's engineers are vendor-neutral, work this way, and leave the eval suite with the team.
Related
AI agent development
Agents that do real work for users, with the evals, traces and guardrails to keep them in production.
DeepSeek vs ChatGPT
DeepSeek's open-weight V4 models and OpenAI's GPT-6 family compared on price, limits, licence and hosting.
DeepSeek vs Claude
DeepSeek's open-weight V4 models and Anthropic's Claude 5.5 models compared for agent builders.
Kimi vs Claude
Moonshot's open-weight Kimi K3 family and Anthropic's Claude models compared for agent builders.
Eval
Repeatable, scored tests that tell a team whether an AI change made things better or worse.
Running agents that have to hold up?
Describe the agent, the systems it touches and what it must never get wrong. The first call is with a lab engineer, who will say plainly whether it is ready for production, and whether Motiontic is the right team.