Analysis2026-11-13Updated 2026-11-13

DeepSeek V4.1 Flash vs Qwen 4 Max Preview: What the Horse Demos Actually Show

TL;DRDeepSeek V4.1 Flash-1017 offers published API rates and MIT-licensed weights, while Qwen 4 Max Preview supports video input during a time-limited Token Plan preview. A community horse experiment illustrates both workflows but does not establish a general coding winner.

By Cameron Voss and West Mao

What happens when both models build the same horse?

In a community experiment, Reddit user migsperez gave the models a detailed brief: create a procedurally modelled, animated low-poly horse in Three.js. The author supplied three reference images, motion references, research material and access to Playwright MCP. After the initial run, each model received one follow-up prompt listing the issues to address.

The author ran DeepSeek V4.1 Flash through OpenCode TUI and Qwen 4 Max Preview through Qoder CLI. They reported that DeepSeek needed fewer self-correction turns during 3D modelling. That is the creator's observation, not a verified measurement from a published run log. We did not run either model ourselves for this article.

What should you look for in a horse animation?

Start with the side view and watch a full stride. Does each hoof move plausibly, or slide across the ground? Then rotate to the front and rear: a convincing silhouette from one angle can conceal awkward joints or symmetry elsewhere. Check whether changes in gait affect the timing and footfall, not just the speed of one repeated loop.

Next, inspect the wireframe and interface if available. The supplied brief asked for 2,500–3,500 horse triangles, separate from the scenery. The DeepSeek result displays a claim of 3,492 horse triangles; treat that as the demo's own label unless you independently instrument the renderer. The Qwen result also exposes a triangle counter. A valid count still says little about anatomy or animation quality by itself.

There is a catch: the original task explicitly prioritized the horse over the interface. An attractive control panel or landscape may make the result more enjoyable, but it should not compensate for weak anatomy when judging whether the brief was met. The published task and reference description make that distinction clear.

Is the horse test a fair model comparison?

It is a useful comparison of two complete workflows, with meaningful limitations. Both got a detailed brief and access to browser automation, but OpenCode and Qoder can differ in system prompts, tool orchestration, context management and how they handle errors. Model settings, run costs, timing and complete transcripts are not established in the post. The one follow-up per run can also contain different feedback for different outputs.

Those differences matter because a coding agent is a loop: the model proposes an edit, tools expose the result, and the agent decides what to do next. A successful run reflects that entire loop. A single creative task, especially one with subjective visual criteria, cannot prove that either model is better at debugging production code, writing tests or maintaining a large repository.

Read the result narrowly: the experiment shows what these particular model-and-tool combinations produced under this author's conditions. It does not measure an isolated model capability or a repeatable win rate.

How do DeepSeek and Qwen compare on specifications?

Three numbers to orient your comparison

Official specifications, checked November 13, 2026

1MContext tokens for each model
Image + videoQwen preview's documented media input; DeepSeek Flash supports images
Nov 25Qwen preview end date, UTC+8
DimensionDeepSeek V4.1 Flash-1017Qwen 4 Max Preview
Official model IDdeepseek-flash on DeepSeek's APIqwen4-max-preview on Alibaba Model Studio
InputsText and images on the official Flash APIText, images and video
Context / maximum output1M tokens / 384K output1M tokens / 393,216 output tokens, shared between thinking and the final answer
ToolsTool calls and JSON output; Search and Fetch are separate APIsFunction calling, built-in tools, web search and structured outputs
DeploymentAPI plus published 1017 weights under MITPreview through Token Plan; the cited documentation does not establish downloadable preview weights
AvailabilityOfficial API and open-weight checkpointPreview ends November 25, 2026 at 24:00 UTC+8; Alibaba says this preview model will be retired afterward

Specifications checked November 13, 2026. Provider implementations can differ; check the relevant API's limits before deploying.

The context windows look equal on paper, but output budgets and supported media differ. If your workflow must inspect video directly, Alibaba documents that capability for Qwen. If you want control over hosting or a published model license, DeepSeek's 1017 checkpoint gives you an option Qwen's preview documentation does not establish. Open weights do not mean the model is lightweight: hardware, inference support and operating cost still need an engineering assessment.

How do the costs compare?

DeepSeek publishes metered API prices; Qwen 4 Max Preview uses Token Plan credits and does not support pay-as-you-go access during the preview. That means a direct per-million-token comparison is not available from the two official pricing schemes. Plan credit consumption, allowances and actual task duration matter more than a headline subscription fee.

DeepSeek official API, per 1M tokensPeakOff-peak
Input, cache miss$0.30$0.15
Input, cache hit$0.006$0.003
Output$1.20$0.60

For illustration, a DeepSeek task using 1M uncached input tokens and 200,000 output tokens would cost $0.54 at peak or $0.27 off-peak in model-token charges. This excludes search, fetch, other tools and any agent retries outside those token totals. Cached input and a different mix of tokens change the result. DeepSeek's published peak schedule and rates can change, so verify them before budgeting.

One DeepSeek task: $0.54 peak vs $0.27 off-peak

Illustration: 1M uncached input tokens + 200K output tokens

Peak
$0.54
Off-peak
$0.27

Model-token charges only. Search, fetch, other tools and additional usage are excluded. Qwen's credit-based preview cannot be directly plotted on this basis.

What do other demos and benchmarks tell us?

DeepSeek's own 1017 announcement shows examples including a web racing game, a 3D Forbidden City and a Godot recreation. These are vendor-selected demonstrations of what can be built, not independent head-to-head results. DeepSeek also publishes benchmark results and identifies the harnesses used for some tests. Qwen 4 Max Preview is not in that model card's comparison table, so the numbers cannot establish a winner between the two.

For another hands-on example, HorseBench lists a separate DeepSeek Flash horse demo alongside its community collection. It is a different run from the Reddit comparison and should be treated as an illustration, not a matched Qwen result. If you want a broader view of Alibaba's early preview, our Qwen launch and early performance report discusses a pelican animation, another 3D scene and a community leaderboard, each with its own limitations.

How can you run a more useful comparison?

Use a task that resembles your work, then control the variables that the horse experiment leaves open. For a frontend task, start with the same empty repository, dependencies, assets and browser viewport. For a maintenance task, use identical copies of a repository with the same failing tests. Keep the agent harness, tool permissions, reasoning settings where comparable, and stop conditions aligned.

  1. Write the acceptance tests first. For a horse, define separate checks for anatomy, gait, triangle budget, console errors and interaction. For application code, use executable tests and a short human review rubric.
  2. Set a fixed budget. Record elapsed time, tool calls, input and output tokens, spend or credit consumption, and number of human interventions.
  3. Preserve the traces. Save prompts, model and provider IDs, harness versions, tool outputs, commits and screenshots. Label every retry and correction.
  4. Repeat and blind-review. Run multiple trials per model and have reviewers score anonymized results. Report failures as well as successes.
  5. Compare the full workflow. Include the time required to review, fix and ship the result. A fast first draft can lose its advantage if the repair bill is large.

A compact prompt for your own test

The following is a new, shorter test prompt inspired by the same task; it is not the author's original prompt. Give both agents the same assets and environment.

Build a browser-based low-poly horse in Three.js without imported 3D models.
Provide idle, walk, trot, canter and gallop controls, orbit camera controls,
and a wireframe toggle. Keep the horse between 2,500 and 3,500 triangles,
excluding the environment, and show the measured horse-only triangle count.
Use the supplied reference images for proportions and gait. Test the page
in a browser, fix console errors, and explain any requirements you could not meet.

For a fairer test, specify in advance whether web access is allowed, which reference files are available, how many correction rounds are permitted and what constitutes completion. Do not supply the models with each other's output.

Which model should you choose?

Choose DeepSeek V4.1 Flash first when you need published metered API pricing, an open-weight path, or a practical baseline for a code agent. Its horse demo and vendor examples make it a strong candidate for evaluation, but neither is a guarantee for your codebase.

Try Qwen 4 Max Preview when native video understanding or its particular Alibaba tooling is relevant and you can work within the preview's access and retirement window. Treat it as an experiment whose integration may need to change after the preview. For a high-stakes workflow, run the same real tasks with both and choose on accepted outcomes and total review effort.

Frequently asked questions

Did DeepSeek beat Qwen in the horse test?

The experiment's author said DeepSeek needed fewer self-corrections for modelling. The post does not provide enough standardized data to establish a general winner, and the two runs used different coding tools.

Can Qwen 4 Max Preview analyze video?

Yes. Alibaba documents image and video input for qwen4-max-preview, subject to the provider's limits and access conditions.

Can DeepSeek V4.1 Flash run on my own infrastructure?

DeepSeek publishes the 1017 model weights under MIT. You still need compatible inference software and sufficient hardware; assess both before planning a self-hosted deployment.

Does 1M context mean both models behave the same on long tasks?

No. A context limit is a capacity specification, not a measure of retrieval accuracy, latency or cost. Test long tasks with your own documents and success criteria.

Sources