DeepSeek vs Qwen: prices, context and benchmarks compared

DeepSeek and Qwen are the two Chinese open-weight model families most developers compare, and the choice comes down to what you are building. This comparison of DeepSeek vs Qwen, as of November 1, 2026, covers DeepSeek V4.1 Flash and Qwen3.8. DeepSeek is cheaper through its API and has the bigger context window. Qwen has a small model you can run on your own hardware under Apache 2.0.

This comparison is based on the products' docs, model cards and published results. We did not run either model ourselves.

DeepSeek vs Qwen at a glance

Prices are from DeepSeek's pricing page and Alibaba Cloud Model Studio (Singapore region), as of November 1, 2026. Both vendors say prices can change.

How much DeepSeek and Qwen cost

DeepSeek charges by time of day. Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays, excluding Chinese public holidays. Off-peak rates are half the peak rates. The larger deepseek-v4-pro costs $0.66 input and $1.98 output per million tokens off-peak, and does not support vision.

Alibaba Cloud has a flat price per model. Qwen3.8 Flash costs about the same as DeepSeek Flash on input ($0.15 against $0.15 to $0.30) and slightly less on output ($0.47 against $0.60 to $1.20). Qwen3.8 Max costs $2 input and $6 output, roughly three times DeepSeek V4 Pro on input. Alibaba lists a free quota of 1 million tokens per model, valid for 90 days.

Context caching changes the picture for long prompts. DeepSeek lists cache-hit input at $0.003 per million tokens off-peak. Alibaba says both Qwen models have a caching discount but the page we read does not give the rate.

Which is stronger: published benchmarks

Each lab publishes scores that favor its own model, so treat these as claims.

• DeepSeek says V4.1 Flash-1017 scores 42.6 on Humanity's Last Exam without tools, 42.3 on Terminal-Bench 4.0 and 77.0 on DeepSWE v1.1. Its card compares it with Kimi K3, GLM-5.3, GPT-5.6 Sol and Claude Opus 5, not with Qwen.
• Qwen says Qwen3.8-27B scores 61.7 on SWE-bench Pro, 73.0 on Terminal Bench 2.1, 84.3 on OSWorld-Verified and 89.2 on GPQA Diamond. Its card compares it with Qwen3.6-27B, Qwen3.7-Plus and Claude Opus 4.6, not with DeepSeek.

The two cards use different benchmark versions and different rivals, so the numbers cannot be compared line by line. Terminal-Bench 4.0 and Terminal Bench 2.1 are not the same test.

Independent numbers from Artificial Analysis

Artificial Analysis runs models through one set of tests. On November 1, 2026, its pages showed these figures. The pages carry no date stamp, so scores may move.

Artificial Analysis lists the DeepSeek model as the top open-weights model on its index. We found no index score for Qwen3.8 Flash or Max on the pages we read, so this table covers only the Qwen model that has one. The two models also differ in size: a 763B model against a 27B one.

Which one to pick

Pick DeepSeek if you call a hosted API and want the lowest price per token, a 1M-token window by default, or image input on the Flash model. The MIT license also lets you download the weights, though 763B parameters is far beyond one consumer GPU.

Pick Qwen if you want to run a model yourself. A 27B dense model is within reach of one workstation GPU once quantized, and Apache 2.0 allows commercial use. Our guide on running Qwen 3.8 locally covers the setup. Qwen also offers Flash and Max through Alibaba Cloud if you need a hosted service.

Both have caveats. DeepSeek's off-peak prices depend on the clock. Qwen's 27B model has no first-party hosted price yet, and its long-context mode can hurt short-text quality, according to its model card.

What we don't know

Neither lab has published a head-to-head against the other. Alibaba's pages do not list parameter counts for Flash or Max. We did not test either model, so we cannot say how they behave on your own tasks.

More on each: DeepSeek V4.1 Flash explained, is Qwen 3.8 open source and DeepSeek vs Claude.