DeepSeek V4.1 Flash Goes Native Multimodal: Vision Trained From Scratch and the 95.6% DocVQA Breakthrough

DeepSeek V4.1 Flash Goes Native Multimodal: Vision Trained From Scratch and the 95.6% DocVQA Breakthrough

For most of 2026, DeepSeek’s vision capabilities lived in a side project. The experimental V4-Flash-Vision-Exp model, released on August 21, 2026, attached a vision encoder to an already-finished text model and delivered promising results on visual agent benchmarks. But it was explicitly labeled experimental, carried a separate API endpoint, and was never meant to be the production answer.

That changed on September 10, 2026, with the release of DeepSeek-V4.1-Flash. For the first time, a DeepSeek model ships with native multimodal vision trained from scratch — images were part of the 45-trillion-token pretraining corpus from day one, and the vision encoder (dubbed DeepSeek-ViT) was trained from scratch rather than borrowed from an existing model. There is no separate vision build, no “Exp” suffix, and no second API endpoint. Image input lives in the main model behind a single identifier: deepseek-flash.

From Bolt-On to Built-In

The difference between V4-Flash-Vision-Exp and V4.1 Flash is not incremental — it is architectural.

V4-Flash-Vision-Exp took the text-only V4 Flash model and attached a vision encoder as an external plugin. Text and images were processed through separate pathways and fused late. This approach works, but it has limitations: cross-modal reasoning is weaker because the two modalities never shared a common representation space during training, and latency suffers from the coordination overhead between separate encoders and decoders.

V4.1 Flash does it the other way around. According to the model card on Hugging Face, the backbone is a 552-billion-parameter mixture-of-experts model pretrained on a corpus that included text and images together from the start. The DeepSeek-ViT vision encoder was trained from scratch alongside the text backbone. The total parameter count reaches 763 billion with the encoder attached, but only 8 billion are active during prefill and 16 billion during decode — which is how a model this large still runs at Flash speed and Flash prices.

The result is a unified latent space where text and images are represented in the same terms. The model can reason across modalities more effectively, and there is no separate encoding pipeline to coordinate.

The Vision Benchmarks

DeepSeek reports four key vision benchmarks for V4.1 Flash:

Benchmark What It Measures V4.1 Flash Score
MMMU-Pro College-level questions requiring both image and text 56.5
CVBench Counting, depth ordering, and spatial relations in natural photos 77.9
DocVQA Question answering over scanned documents and forms 95.6
RefCOCO Locating the object a phrase refers to inside an image 86.0

Two of these scores are particularly consequential for agent applications.

DocVQA at 95.6 means V4.1 Flash can extract information from scanned documents and forms with near-perfect accuracy. For enterprise workflows — invoice processing, form intake, contract review, receipt parsing — this is the capability that turns paper and screenshots into structured data. A 95.6% score means the model gets the vast majority of fields right on the first pass, dramatically reducing the human review burden.

RefCOCO at 86.0 measures grounding — the ability to locate a specific object described in natural language within an image. Given the instruction “click the Submit button below the email field,” a model with strong RefCOCO performance can find the right UI element in a screenshot. This is the foundational capability for screenshot-driven agents: agents that observe a screen, understand what they see, and take precise actions. RefCOCO 86.0 is strong enough to make screen-based automation genuinely reliable.

How It Works in Practice

The API experience for vision is straightforward. V4.1 Flash accepts images in three formats through the standard Chat Completions endpoint at https://api.deepseek.com:

  1. Base64 data URL — up to 32 MiB per image
  2. External URL — up to 8,192 characters
  3. File ID — for reusable image assets

An optional detail field accepts low, high (alias original), or auto, letting developers trade off between speed and visual fidelity depending on the task. Context length is 1 million tokens, with a maximum output of 384K tokens — the same as text-only calls.

Pricing is the standard Flash rate: $0.15 per 1 million cache-miss input tokens off-peak and $0.30 peak, with output tokens billed separately. This means multimodal capability is available at the same price as text — a significant departure from providers that charge a premium for image input.

The Retirement of V4 Pro

The most aggressive signal in the V4.1 Flash release is not a benchmark number — it is a routing decision. DeepSeek announced that after 12:00 Beijing Time on September 14, 2026, all API requests to deepseek-v4-pro would be routed to V4.1 Flash and billed at the V4.1 Flash price, at least until a future V4.1 Pro model is released.

The reason given: extensive internal testing showed V4.1 Flash outperforms V4 Pro across performance, cost, speed, and total time. When your entry-tier model beats your flagship on every dimension, the flagship becomes redundant. The previous-generation models V4 Flash and V4 Flash Vision Exp were also retired, with their model names temporarily routed to V4.1 Flash for backward compatibility.

This is a rare move in the AI industry — most providers keep a clear performance gap between their Flash and Pro tiers to justify tiered pricing. DeepSeek’s decision to collapse that gap suggests the new architecture family delivers efficiency gains large enough to make the old pricing ladder obsolete.

Speed That Enables Real-Time Interaction

Native multimodal would be less impressive if it came with a latency penalty, but V4.1 Flash is faster than its predecessor across the board. Community benchmarks report peak performance of 420 tokens per second on long-text reasoning tasks, with speedups ranging from 3.9x to 6.0x over V4 Flash on tasks including:

  • 49K long-context retrieval: 5.2x faster
  • SVG code generation: 6.0x faster
  • Manacher palindrome algorithm: 4.6x faster
  • Complex SQL query generation: 5.0x faster
  • Asyncio architecture refactoring: 3.9x faster

For multimodal interaction, this speed matters. A user who uploads a screenshot and asks “what’s wrong with this UI?” expects a response in under a second, not ten. V4.1 Flash’s throughput makes real-time multimodal conversation practical.

What This Means for the DeepThink Ecosystem

For the DeepThink community, V4.1 Flash’s native multimodal capability completes a picture that has been forming since the release of the V4 model family. The reasoning traces that DeepThink pioneered — long, auditable chains of thought — are now embedded in a model that can also see, read documents, locate UI elements, and reason across text and image simultaneously.

Three implications stand out:

  1. Document-heavy workflows become agent-ready. DocVQA at 95.6% means agents can process invoices, forms, contracts, and receipts with minimal human oversight. This unlocks a broad class of enterprise automation that was previously bottlenecked on OCR accuracy.

  2. Screenshot-driven agents become reliable. RefCOCO at 86.0 gives agents the grounding capability to interact with graphical interfaces by observing screenshots. Combined with Terminal-Bench 2.1 scores of 90.6, this creates agents that can operate both in the terminal and on the desktop.

  3. Multimodal stops being a premium feature. By bundling native vision into the Flash tier at no extra cost, DeepSeek is making multimodal the default rather than the upgrade. This forces the broader market to reprice vision capabilities and accelerates adoption among cost-sensitive developers.

Conclusion

DeepSeek V4.1 Flash’s native multimodal design is more than a feature upgrade — it is an architectural statement. By training vision from scratch alongside text, by unifying both modalities in a single model and a single API, and by delivering it at Flash speed and Flash prices, DeepSeek has made multimodal reasoning the baseline rather than the bonus.

The 95.6% DocVQA score and 86.0% RefCOCO score are not just numbers on a benchmark sheet. They represent the point at which document understanding and screen grounding become reliable enough to power production agent workflows. When combined with the model’s elite coding ability (Codeforces 3,471) and class-leading speed (420 tokens/second), V4.1 Flash emerges as a genuinely general-purpose agent platform.

For the DeepThink ecosystem, the message is clear: the reasoning engine now has eyes. The agents built on top of it can read documents, understand screenshots, and act on what they see — all through a single, fast, affordable API.