DeepSeek V4.1 Flash Hands-On: Does It Really Beat V4 Pro Across the Board?

DeepSeek V4.1 Flash Released: How Did a Flash Model Overtake the Flagship?

On September 10, 2025, DeepSeek officially released the V4.1 Flash model. The most intriguing aspect of this launch: before it even went live, DeepSeek explicitly stated that V4.1 Flash comprehensively outperforms V4 Pro — and that all future requests to V4 Pro would be automatically routed to V4.1 Flash.

This defies conventional expectations. The Flash series has always been positioned as a lightweight model prioritizing speed and cost-efficiency, while V4 Pro was DeepSeek's flagship. The fact that a cost-oriented Flash model is now replacing the flagship signals a genuine architectural leap — not just a cost-cutting rebrand.

This article draws on real-world test data to break down exactly what V4.1 Flash improves, and how it performs in actual coding scenarios.

Two Core Upgrades: New Architecture and Native Multimodal Support

All-New Model Architecture with a Higher Capability Ceiling

V4.1 Flash is built on an entirely new model architecture, delivering three direct benefits: a higher capability ceiling, faster inference speed, and greater throughput. This explains how a Flash-tier model can surpass the old flagship — architectural iteration often drives more meaningful capability jumps than simply stacking more parameters.

For DeepSeek, replacing the aging V4 Pro with a more efficient new-architecture model improves overall service quality while lowering inference costs — a classic "technology upgrade that also optimizes cost structure" move.

Native Multimodal Vision Understanding

This is the most significant addition in this release. Previously, using multimodal capabilities on DeepSeek required calling specialized experimental models like V4 Flash v-simple. Now, V4.1 Flash natively supports image understanding, covering use cases such as screenshot OCR, image description and content comprehension, chart analysis, and webpage screenshot interpretation.

The practical value of multimodal capabilities is often underappreciated. When you encounter a bug or error, describing the problem in text is often awkward and imprecise — but if you can just throw a screenshot at the model and let it diagnose the issue visually, the efficiency gain is substantial. Native multimodal support transforms V4.1 Flash from a pure code/text model into something that fits much more naturally into real development workflows.

Pricing Changes: Time-of-Day Billing

Starting at noon on September 10, V4.1 Flash operates under a new pricing structure with time-of-day tiered billing:

• Off-peak hours: ~¥1 per million input tokens, ~¥4 per million output tokens
• Peak hours: Prices double

This tiered pricing strategy essentially uses price signals to encourage users to schedule non-urgent, high-volume tasks during off-peak hours, smoothing out server load. Developers running batch processing or offline tasks can meaningfully cut costs by timing their workloads accordingly.

Benchmark 1: Web-Based Excel Generator

The first test task was building a browser-based Excel application, compared side-by-side with the previous V4 Flash. Testing used DeepSeek's official Harness (not CodeX). One useful tip worth noting:

npx deepseek  # Pin the version number — otherwise the new model may not appear in the UI

If V4.1 Flash doesn't show up in the model selector, just specify the version number directly when running the command.

Here are the results:

Model	Time	Cost	Test Period
V4 Flash	9 min 47 sec	¥0.17	Off-peak
V4.1 Flash	21 min	¥3.68	Peak

V4.1 Flash produced a far more feature-complete web Excel app — including file editing, insert operations, formatting options, and formula demonstrations. That said, it's worth noting that this comparison has two confounding variables: the tests ran at different times of day (off-peak vs. peak, meaning double the price), and V4.1 Flash tackled a more complex feature set. These numbers are best interpreted as "more capable but more expensive," not as a raw performance regression.

Benchmark 2: Website Replication from Screenshot (Multimodal + Frontend Reconstruction)

The second test was replicating a website from a screenshot, designed to stress-test both native multimodal understanding and frontend reconstruction ability. The prompt was: "Replicate the website shown in the image as faithfully as possible."

Results:

• V4 Flash v-simple: Recognized the basic page structure but missed details in the middle section. Completed in 2 min 54 sec, costing ¥0.10. Extremely fast.
• V4.1 Flash: More complete recognition — header, body, and footer all reconstructed with reasonable fidelity to the original. Completed in 8 min 26 sec, costing ¥1.40.

V4.1 Flash shows a clear improvement in multimodal visual accuracy, particularly in capturing finer page details. The trade-off is increased time and cost — the old model wins on speed and price, the new model wins on accuracy and completeness.

Verdict: Worth Upgrading, Depending on Your Use Case

The core logic behind DeepSeek V4.1 Flash is straightforward: replace the aging V4 Pro with a more capable new-architecture model, while finally filling the gap with native multimodal support. From real-world testing:

• More capable across the board: Both the web Excel generator and website replication tasks produced more complete, more accurate results with V4.1 Flash.
• Native multimodal is the standout feature: No more experimental vision models — just pass in an image and go. This meaningfully improves debugging and UI reconstruction workflows.
• Cost requires consideration: V4.1 Flash takes longer and costs more per run. Combined with the peak/off-peak pricing tiers, scheduling heavy workloads during off-peak hours is strongly recommended.

For simple, high-frequency tasks where raw speed and low cost are the priority, the older Flash model may still hold an edge. But if you need stronger overall capability and native multimodal support, V4.1 Flash is the clear upgrade. As always, the most reliable way to know if it works for you is to run your own tests.