DeepSeek V4.1 Flash vs Qwen 4 Max Preview
TL;DRDeepSeek V4.1 Flash is the more practical default for a production API or self-hosted evaluation today: its 1017 weights are available under MIT and its official API has published token prices. Qwen 4 Max Preview is worth testing for native video input and its coding or agent workflows, but its QwenCloud API is a time-limited Token Plan preview. We have not found a controlled head-to-head test that establishes which model is more accurate.
By Cameron Voss and West Mao
At a glance
| Dimension | DeepSeek V4.1 Flash | Qwen 4 Max Preview |
|---|---|---|
| Access | Official API: deepseek-flash, currently serving V4.1 Flash-1017; downloadable weights | qwen4-max-preview on QwenCloud Token Plan; preview ends November 25, 2026, 24:00 UTC+8 |
| Input | Text and image on the official Flash API | Text, image and video; text output |
| Context / output | 1M context; up to 384K output | 1M context; up to 393,216 output tokens, including thinking |
| Tools | Tool calls and JSON output; Search and Fetch are separate billed APIs | Function calling, built-in tools, web search and JSON Schema |
| Price basis | Metered API input, cached input and output, with peak and off-peak rates | Token Plan credits; no pay-as-you-go rate for the preview |
| Weights | 1017 checkpoint published under MIT | Preview weights and license not established in the cited documentation |
Explore your options: Explore the DeepSeek subscription and current Coding Plan options. Review the current availability, allowance and terms before choosing.
Coding and agent performance: what is actually known?
Both vendors target coding and tool-assisted work, but that does not settle a winner. Alibaba describes Qwen 4 Max Preview as suited to complex programming and frontend work; its current model documentation does not supply a matched benchmark against Flash. DeepSeek publishes a model and research materials, but a result for Flash alone cannot be turned into a head-to-head score without a comparable Qwen run.
For a coding agent, separate a patch that compiles from a patch that solves the issue. A useful test asks each model to inspect the same repository, reproduce a failing test, modify several files and explain remaining risks. Measure passing tests, retries, tool errors and review time. A frontend demo also needs a visual review: a page can render and still have weak spacing, accessibility or mobile behavior.
Images, video and long documents
Both official APIs list image input, useful for screenshots, diagrams and scanned pages. QwenCloud also documents video input for the preview. If the work starts with a recording, that is a concrete feature difference; check provider limits and sample rates for your particular request. Do not infer that DeepSeek's API accepts a video file merely because its model accepts images.
Both advertise a 1M context window. Qwen's 393,216-token maximum includes thinking and final output; DeepSeek lists a maximum of 384K. These limits describe different API contracts, so reserve room for the prompt, tool responses and reasoning before choosing output limits. Test retrieval at the length you will actually use.
API price: a calculation for Flash, not a false Qwen comparison
DeepSeek's official Flash rates per million tokens are $0.15 uncached input and $0.60 output off-peak; peak rates are $0.30 and $1.20. Cached input is $0.003 off-peak or $0.006 peak. Peak is 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays, excluding Chinese public holidays.
For an illustrative job with 1 million uncached input tokens and 200,000 output tokens, the model charge is $0.27 off-peak or $0.54 peak, before any separate Search or Fetch calls. Qwen's preview uses subscription credits rather than a public pay-as-you-go rate, so a dollar-per-million comparison would be misleading. Check the Token Plan price, credit deduction, quota, region and permitted usage in your account; a subscription's effective cost depends on how much of its allowance you consume.
Integration, continuity and deployment
DeepSeek documents OpenAI-compatible and Anthropic-format endpoints, tool calls and JSON output. QwenCloud documents OpenAI Chat Completions, Responses and Anthropic Messages compatibility, with a dedicated Token Plan key and base URL. Protocol compatibility does not guarantee identical parameters, tool behavior or errors. Test the exact client and provider rather than only changing the model name.
QwenCloud says the preview will be retired after November 25 and says it will not change during the stated preview period. That commitment applies to the documented preview period; it does not establish a permanent replacement. Earlier anonymous Qoder tests were reported as changing daily before the named preview. Keep a fallback and do not assume an existing model ID will remain available. Flash's downloadable checkpoint gives another deployment option, though 510 GB of repository files makes self-hosting a substantial infrastructure decision.
Which should you choose?
- Choose Flash first for metered API workloads, a documented production route, or a self-hosting investigation where MIT terms matter.
- Try Qwen first when direct video input or its preview coding workflow matters and the Token Plan terms fit your use.
- Keep both in evaluation for high-value coding tasks until repeated, matched tests show which produces fewer corrections for your own repository.
What changes when you use an agent instead of chat?
A chat answer is one response to one prompt. A coding agent may search a repository, open files, run tests, edit code and repeat that cycle many times. At each step, the model must decide what to inspect next and interpret the result. A model that gives a convincing explanation can still struggle when a tool returns an unexpected error.
Start with a bug that has a reproducible failure and a test that exposes it. Give each agent the same checkout and permission to run the same commands. A useful result identifies the cause, changes the smallest relevant surface and passes both the new test and existing tests. Record unnecessary file edits and whether the model declares success before checking its work. Those details are often more useful than a subjective preference for one answer.
For a longer task, ask both agents to implement a feature across the UI, API and persistence layer. Check the contract between layers, error handling and rollback behavior. If the agent needs a second prompt, record what the human had to explain. A lower-cost model can be the better choice for routine work, but only if its extra corrections do not absorb the savings.
Frontend work: code quality and visual quality are separate
Both models can be asked to build a landing page or reproduce a screenshot. Judge the result in a browser at more than one width. Does the layout match the brief? Do buttons work? Can a keyboard user reach the controls? Does text remain readable when the viewport narrows? A visually attractive screenshot is not enough if navigation or form validation fails.
Qwen's video input may help when a design brief includes a screen recording rather than still images. Flash's image input can still be useful for screenshots and diagrams. The sensible test is to provide the same still frames to both, then run a separate Qwen video test. That separates model quality from the extra information available in a video.
Long context: capacity is not the same as useful recall
Consider a repository with a recent authentication change and an older design document that contradicts it. Put both documents in the supplied context and ask for a patch. A strong answer should identify which evidence is current, use the actual code as the source of truth and explain any unresolved conflict. Repeat the task with the relevant information near the start, middle and end of the context.
Large prompts also cost time and money. A one-million-token window lets a team submit more material, but it does not remove the need to select relevant files. An agent with search and file tools may do better by opening a smaller set of files, provided it can find the right ones. Compare total input and cache usage across the whole session, not just the initial prompt.
Privacy and deployment choices
A hosted API sends the supplied material to the chosen provider. Review that provider's retention, regional processing and contractual terms before sending private code or customer data. Protocol compatibility does not make two providers' data policies identical. If a team uses a third-party gateway, its terms and logging practices matter too.
DeepSeek's published weights make private deployment possible in principle, but the size of the checkpoint and its serving requirements make this a serious infrastructure project. Include accelerator capacity, memory, operations, monitoring and an appropriate security review in the decision. Qwen's documented preview is a hosted service; the material reviewed does not establish downloadable preview weights. Neither option should be described as private by default.
What to do before the preview ends
Run a small evaluation while access is available and keep the results in a format your team can repeat. Save prompts, input versions, settings, outputs, scores and the date of each run. Avoid sending confidential material until the relevant provider terms have been reviewed. If the preview performs well, keep a fallback model and repeat the evaluation when a successor becomes available. An improvement in one preview build does not guarantee the same behavior from a later model.
How to run a fair comparison
- Choose 20 representative tasks before looking at results: bug fixes, multi-file changes, screenshot-to-code work, and questions that require citations. Include cases where your current model fails.
- Use the same repository snapshot, acceptance tests, tools, time limit and human review criteria. Record model ID, provider, region, date, reasoning settings, context and agent version. The preview may have no user-visible build identifier.
- Run each task several times. Score the fraction that passes hidden tests, correctness of tool calls, security regressions and human correction time. Keep failed runs in the results.
- Record time to first token, output tokens per second and end-to-end completion time separately. Count input, cached input, output, retries and any search charges. A fast stream is not the same thing as a fast, correct solution.
- For long-context work, place known facts at different positions and check retrieval and contradictions at several lengths. A 1M context limit is a capacity claim, not proof of reliable recall at 1M.
Frequently asked questions
Is Qwen 4 Max Preview better at coding?
There is not enough matched evidence here to rank it against Flash. Compare both with the same repository, agent and tests.
Can I use Qwen 4 Max Preview on a pay-as-you-go API?
QwenCloud says the preview is Token Plan only. Check the plan's current terms and availability for your region.
Is DeepSeek V4.1 Flash open weight?
Yes. DeepSeek publishes the 1017 checkpoint under MIT; hosting it requires significant resources.