DeepSeek vs Qwen in 2026: Pricing, Models and Licensing Compared

If you have a budget spreadsheet open and you’re trying to choose between two Chinese model families that both publish open weights, the deepseek vs qwen decision usually comes down to four numbers: input price, output price, context length, and how the model behaves on your actual workload. Both families publish open weights, both expose OpenAI-compatible APIs, and both also offer Anthropic-compatible endpoints. They are not, however, interchangeable. This article walks through pricing, vendor-reported benchmarks, context windows, licensing, and a worked cost example, using each vendor’s own documentation as of October 2026.

Verdict: which one should you pick?

For most developers building chat, RAG, or agent workloads, DeepSeek’s deepseek-flash is the simpler choice, and on the worked example below it is also the cheaper one against Qwen3.7-Plus at list price. It costs $0.30 per million uncached input tokens and $1.20 per million output tokens at peak, half that off-peak ($0.15 / $0.60), with a 1M-token context window. Qwen’s hosted lineup on Alibaba Cloud Model Studio is broader — Qwen3.8-Max at the top, Qwen3.7-Plus for balanced workloads and Qwen3.8-Flash for speed-first tasks — and Qwen3.8-Flash’s list price ($0.15 input / $0.47 output) is close to DeepSeek’s. But Qwen’s pricing is spread across tiers, regions and context-length bands, and its open-weight releases come under several different licences.

Pick DeepSeek if you want two API models on one pricing page, a single peak/off-peak rule, and MIT-licensed weights. Pick Qwen if you need hosted video or audio models, a choice of deployment regions, Alibaba’s Qwen Code agent tooling, or a smaller open model (for example the 27B dense Qwen3.8-27B) you can run on far less hardware.

At a glance: DeepSeek vs Qwen in one table

The headline tiers each family lists as of October 2026. DeepSeek prices are from DeepSeek’s Models & Pricing page (peak / off-peak); Qwen prices are Alibaba Cloud Model Studio standard list prices for the Singapore region, International deployment scope, smallest input-length tier (page last updated 22 October 2026). Qwen weights and licences are from the Qwen organisation on Hugging Face.

DeepSeek’s off-peak rates apply outside 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday (Chinese public holidays and weekends are off-peak in full). Model Studio lists Qwen3.7-Plus at a higher rate for requests above 256K input tokens ($1.20 input / $4.80 output) and showed a limited-time 20% discount on it; the older Qwen3.6-Plus is still listed at $0.50 / $3. Qwen rates differ by region, so always check the provider’s live pricing page before committing.

Architecture and lineage

Both families use Mixture-of-Experts (MoE) at the top end, and Qwen also ships dense models at the smaller sizes, but they get there differently.

DeepSeek: two API models, one pricing page

DeepSeek’s API lists exactly two model IDs. deepseek-flash is served by DeepSeek-V4.1-Flash-1017 (released 17 October 2026) and is the only one with vision; deepseek-v4-pro is served by DeepSeek-V4-Pro-0813, the 13 August 2026 official release that replaced the April preview. Both publish weights under MIT, share a 1M-token context and 384K maximum output, and support thinking and non-thinking modes, with thinking on by default. Thinking depth is a request parameter, not a separate model: reasoning_effort takes "low", "high" or "max". The legacy deepseek-chat and deepseek-reasoner IDs were scheduled for retirement on 24 July 2026 at 15:59 UTC and are no longer listed; the older deepseek-v4-flash name is still accepted but is served by DeepSeek-V4.1-Flash-1017. If you maintain an older integration, the migration is a one-line model= swap; base_url stays the same. For more depth on the family see DeepSeek V4.

Qwen: a wider lineup, more fragmentation

Alibaba’s lineup is broader. Model Studio’s recommended text models (page updated 11 October 2026) are Qwen3.8-Max, Qwen3.7-Plus and Qwen3.8-Flash, alongside omni, image, video, speech and embedding models — and third-party models, including DeepSeek’s own V4-Pro-0813 and V4.1-Flash. On the open-weight side, the Qwen3.8 releases on Hugging Face (August 2026) include a 2.4T-total / 95B-active MoE (Qwen3.8-2.4T-A95B), which its model card describes as the base of the hosted Qwen3.8-Max; Qwen3.8-Flash-Next (125B with 6B activated), the base of Qwen3.8-Flash; and the dense Qwen3.8-27B. Earlier open releases such as Qwen3.6-35B-A3B (April 2026) remain available.

The pattern matters: DeepSeek publishes two API models and prices them on one page. Qwen ships a much wider catalogue, with region- and tier-specific prices and different licences for different checkpoints. If you want a clean mental model and predictable pricing, DeepSeek is simpler.

Coding

This is where both families have invested heavily, and both publish only self-reported scores. DeepSeek’s 13 August 2026 Change Log reports 87.9 on Terminal Bench 2.1 and 62.7 on DeepSWE for DeepSeek-V4-Pro-0813; its 17 October 2026 Change Log reports 77.0 on DeepSWE v1.1 for DeepSeek-V4.1-Flash-1017. Alibaba’s Qwen3.8-2.4T-A95B model card reports 86.6 on Terminal Bench 2.1, 67.7 on SWE-bench Pro and 56.6 on DeepSWE 1.1 for Qwen3.8-Max. The two vendors use different harnesses and settings, so treat these as each vendor’s own claims rather than a head-to-head result.

Neither will save you from a bad spec, and the only comparison that matters is one on your own repository. For a deeper coding-only comparison, see DeepSeek Coder vs Copilot and our DeepSeek for coding walkthrough.

Reasoning

Both families expose reasoning as a parameter rather than a separate SKU. On DeepSeek, thinking is on by default; you set reasoning_effort to "low", "high" or "max", and the API returns reasoning_content alongside the final content. On Qwen, hybrid models are switched with the enable_thinking parameter (passed via extra_body in the OpenAI Python SDK). Alibaba’s deep-thinking docs (updated 19 October 2026) note one quirk: if the model does not output a thought process in thinking mode, the request is billed at the non-thinking price.

Qwen also offers preserve_thinking, which feeds earlier turns’ reasoning_content back into the next round’s input (and bills it as input); Alibaba lists it for the Qwen3.8-Max, Qwen3.8-Flash, Qwen3.7 and Qwen3.6-Plus models, among others. DeepSeek’s rule depends on tools: when a thinking-mode request carries tools, you must pass back the reasoning_content of all previous turns, or the API returns a 400 error; without tools, earlier reasoning is ignored. Either way the API is stateless — the client must resend conversation history with every request. The web chat and mobile app maintain session state for users; the API does not.

Writing and multilingual

Qwen publishes a language count; DeepSeek does not. Alibaba’s Qwen3.5 model card (February 2026) states support for 201 languages and dialects. DeepSeek’s model cards and docs do not publish a comparable figure, so if you need a less widely spoken language, run a sample of your own content through both before you commit.

For long-form English and creative writing, the difference is mostly stylistic. DeepSeek’s official temperature guidance suggests 1.5 for creative writing, but that setting only applies in non-thinking mode — thinking mode, the default, ignores temperature. Try both on a sample of your own copy. See DeepSeek for writing for prompt patterns.

Pricing: a worked example

Comparing token rates across providers is messy. Let’s cost out one realistic workload — a customer-support chatbot doing 1,000,000 calls with a 2,000-token system prompt (cached), a 200-token user message, and a 300-token reply — on each side, using the list prices in the table above (as of 22–23 October 2026).

DeepSeek deepseek-flash

                                                PEAK                     OFF-PEAK
Cached input  : 2,000 × 1,000,000 = 2,000,000,000 × $0.006/M = $ 12.00   × $0.003/M = $  6.00
Uncached input:   200 × 1,000,000 =   200,000,000 × $0.30/M  = $ 60.00   × $0.15/M  = $ 30.00
Output        :   300 × 1,000,000 =   300,000,000 × $1.20/M  = $360.00   × $0.60/M  = $180.00
                                                               -------              -------
Total                                                          $432.00              $216.00

DeepSeek V4-Pro

                                     PEAK                        OFF-PEAK
Cached input  : 2,000,000,000 × $0.044/M = $   88.00   × $0.022/M = $ 44.00
Uncached input:   200,000,000 × $1.32/M  = $  264.00   × $0.66/M  = $132.00
Output        :   300,000,000 × $3.96/M  = $1,188.00   × $1.98/M  = $594.00
                                           ---------              -------
Total                                      $1,540.00              $770.00

Qwen3.7-Plus and Qwen3.8-Flash (Model Studio list prices)

Model Studio bills Qwen3.7-Plus explicit-cache hits at 10% of the standard input price (cache creation at 125%, with a 5-minute validity that resets on each hit). The second Qwen3.7-Plus line assumes the 2,000-token prefix always hits an explicit cache and ignores the one-off creation charge. Alibaba says Qwen3.8-Flash’s cache ratios vary by region, so it is shown without caching:

Qwen3.7-Plus, no caching
Input         : 2,200,000,000 × $0.40/M = $  880.00
Output        :   300,000,000 × $1.60/M = $  480.00
                                          ---------
Total                                     $1,360.00

Qwen3.7-Plus, explicit cache hits on the 2,000-token prefix
Cached input  : 2,000,000,000 × $0.04/M = $   80.00
Uncached input:   200,000,000 × $0.40/M = $   80.00
Output        :   300,000,000 × $1.60/M = $  480.00
                                          ---------
Total                                     $  640.00

Qwen3.8-Flash, no caching
Input         : 2,200,000,000 × $0.15/M = $  330.00
Output        :   300,000,000 × $0.47/M = $  141.00
                                          ---------
Total                                     $  471.00

For this workload deepseek-flash costs $432.00 if run entirely in peak hours and $216.00 entirely off-peak, against $640.00–$1,360.00 for Qwen3.7-Plus at list price and $471.00 for Qwen3.8-Flash before any caching discount. deepseek-v4-pro costs more than Qwen3.7-Plus at peak on this profile ($1,540.00 against $640.00–$1,360.00); off-peak, its $770.00 falls between the cached and uncached Qwen3.7-Plus totals. Promotions (such as the limited-time Qwen3.7-Plus discount) and regional prices change the picture, so verify before you build. Use the DeepSeek pricing calculator to plug in your own token volumes.

Privacy and licensing

Where your data goes depends on the endpoint. DeepSeek’s own API docs do not offer a choice of deployment region, and requests are processed by DeepSeek on servers subject to Chinese law. Alibaba Cloud Model Studio (regions page updated 13 October 2026) lets you pick a region — China (Beijing), Singapore, US (Virginia), Germany (Frankfurt), Japan (Tokyo) or China (Hong Kong) — which sets where request data is stored, plus a deployment scope that sets where inference runs (for example, Virginia supports Global and US scopes, Frankfurt supports Global and EU). Model Studio also hosts DeepSeek’s models: its pricing page lists deepseek-v4.1-flash-1017 and deepseek-v4-pro-0813 in Singapore at the same input (cache-miss) and output rates as DeepSeek’s API, but with Alibaba’s own busy hours (08:00–22:00 UTC+8). If data residency matters, that is a Qwen-platform win — or a reason to self-host DeepSeek’s open weights.

Licensing: every recent DeepSeek release (V4.1-Flash-1017, V4-Pro-0813, V4.1-Flash, V3.2, V3.1, R1) publishes both code and weights under MIT. Qwen is more mixed. Qwen3.8-27B and Qwen3.6-35B-A3B are Apache 2.0, while Qwen3.8-2.4T-A95B ships under the Qwen3.8-Max License and Qwen3.8-Flash-Next under the Qwen Community License 1.0 — both permissive in general, but both add conditions for very large commercial products (prominent model-name display above 100 million monthly active users or US$20 million monthly revenue) and for model-as-a-service businesses, which need a separate licence from Qwen (for Qwen3.8-Max, above US$50 million in twelve-month revenue). Read the licence on the specific Hugging Face repo before deploying — the answer is not always “Apache 2.0”. For more on this see is DeepSeek open source.

Ecosystem and tooling

Both APIs are OpenAI-compatible, so you can switch by changing base_url and api_key. Both also offer Anthropic-compatible endpoints: DeepSeek at https://api.deepseek.com/anthropic, and Model Studio on its regional endpoints. A minimal DeepSeek call with the OpenAI Python SDK — chat requests hit POST /chat/completions:

import os
from openai import OpenAI

client = OpenAI(
    api_key=os.environ.get("DEEPSEEK_API_KEY"),
    base_url="https://api.deepseek.com",
)
resp = client.chat.completions.create(
    model="deepseek-flash",
    messages=[{"role": "user", "content": "Summarise this PR."}],
    reasoning_effort="high",
    extra_body={"thinking": {"type": "enabled"}},
)
print(resp.choices[0].message.content)

Qwen leans further into its own agent tooling: Alibaba maintains Qwen Code, an Apache-2.0, open-source coding agent for the terminal whose README lists MCP support, sub-agents and IDE integrations. DeepSeek’s approach is to plug into existing tools — its docs list DeepSeek Harness (in developer preview) and integrations with agents such as Claude Code and OpenCode, and the API supports tool calls in thinking mode. If you want a first-party CLI agent built around one model family, Qwen has a head start. If you’re building a chat product, retrieval pipeline, or a workflow inside an existing agent tool, DeepSeek’s two-model API keeps things simple.

When to pick DeepSeek vs Qwen

Use this as a decision filter:

• Pick DeepSeek deepseek-flash when you want low per-token cost (with off-peak at half price), a single 1M-token context tier, image input, and MIT-licensed weights you can self-host.
• Pick DeepSeek deepseek-v4-pro when a workload justifies its 3.3× output-price step up from Flash; note that DeepSeek’s 10 September 2026 announcement said tests by multiple parties put V4.1-Flash ahead of V4-Pro on performance, cost and speed.
• Pick Qwen3.8-Max or Qwen3.7-Plus when you need Alibaba’s hosted multimodal lineup, regional deployment scopes, or first-party agent tooling like Qwen Code.
• Pick Qwen3.8-27B or Qwen3.6-35B-A3B when you want an Apache-2.0 model far smaller than DeepSeek’s current open weights.
• Pick Qwen for documented multilingual breadth beyond English and Chinese, given its published 201-language coverage.

Alternatives worth comparing

If you’re shortlisting open and semi-open Chinese frontier models, also look at DeepSeek vs Kimi and DeepSeek vs GLM. For a head-to-head against Western frontier offerings, see DeepSeek vs Claude and the broader AI comparison hub. The best Chinese AI models roundup ranks the full field by use case.