DeepSeek Adds Vision to V4-Flash, Bringing Multimodal Agent Scores Near Opus 4.8

DeepSeek released its first vision-capable model in the V4 family on August 21, 2026, adding image understanding to the text-only V4-Flash model that has been the backbone of its developer ecosystem since April. The new model — identified in the API as deepseek-v4-flash-vision-exp — accepts images alongside text through the same endpoint developers already use for V4-Flash, bills those images at the existing V4-Flash input rate, and carries no vision surcharge. For teams that have been routing images through a separate vision model before passing extracted text to DeepSeek, the wait for a native multimodal option is over.

What the release does not deliver is an independently verified performance claim. The benchmark table that generated most of the coverage — comparing V4-Flash-Vision-Exp against Anthropic's Claude Opus 4.8 — was published by DeepSeek inside a PNG image in the release note, is entirely self-reported, and shows the new model trailing Opus 4.8 on eight of the eleven benchmarks in the table, with the largest gap running to twelve points on the NL2Repo coding task. No third-party evaluator has independently reproduced any of the eleven scores as of publication. DeepSeek's own phrasing — "close to Opus 4.8" — is a reasonable summary of the multimodal-specific rows; the broader coverage framing of "parity" or "challenge" is not supported by the numbers.

The distinction matters because the story worth telling here is not primarily competitive positioning. It is the fact that DeepSeek has eliminated a meaningful architectural dependency for agent developers, and has done so at a price point that makes per-image cost nearly negligible relative to the output tokens those images will generate.

DeepSeek V4 Was Explicitly Text-Only Until This Week

When DeepSeek released V4-Flash and V4-Pro in April 2026, both models were documented as text and code only. Developers building agents on the V4 platform who needed to process images — screenshots of web interfaces, charts, documents, receipts, UI states — had two options: route the image through a separate vision model (typically Gemini or Qwen-VL), extract the relevant text or description, and pass that text to DeepSeek; or accept that V4-based agents simply could not see. The first approach added latency, a second API dependency, and an additional cost layer. DeepSeek Harness, the agent framework the company launched alongside V4, lacked native vision support until the v0.1.1 release that shipped simultaneously with the new model.

V4-Flash-Vision-Exp removes that dependency. It is not a separate model in a new product line — it is an experimental variant of V4-Flash-0731, the retrained July build that currently serves all API calls to the deepseek-v4-flash endpoint. The release note describes it as matching V4-Flash on text capabilities including agents, reasoning, and world knowledge. Whether that matching is literal or approximate is a company claim; the available benchmark data, also company-reported, shows the vision model outperforming its text-only baseline on eight of nine comparable benchmarks, trailing only on Cybergym, which evaluates software vulnerability discovery.

A 284B Sparse MoE Model That Processes Images at a Fixed Token Ceiling

V4-Flash's foundational architecture is a sparse Mixture-of-Experts design with 284 billion total parameters and approximately 13 billion active parameters per token. The MoE approach routes each incoming token through a specialized subset of parameter blocks rather than running the full model, which is why a 284 billion parameter model can be served at inference costs comparable to a much smaller dense network. That same architecture underlies the vision extension: V4-Flash-Vision-Exp inherits the full V4-Flash-0731 text and reasoning stack and adds an image pathway on top.

DeepSeek has disclosed no information about how it built the vision encoder — no paper, no model card section, no architecture description. Third-party Hugging Face repositories with names resembling the official model exist, but their own documentation makes clear they are independent research experiments that splice a MoonViT encoder taken from a competitor's architecture onto the frozen V4-Flash text backbone. As confirmed by the OpenRouter listing, the model has no official Hugging Face repository. These third-party repositories are not official DeepSeek releases and describe nothing about how DeepSeek's production system was constructed.

What DeepSeek has published is how it bills image input. Each image is tokenized internally at up to 384 tokens, regardless of the image's resolution above a certain threshold. The company's Vision guide is explicit on the consequence: a 2,000 × 2,000 pixel image costs identically to a 5,000 × 5,000 pixel image. The guide also specifies supported image formats directly: JPEG, PNG, GIF, and WebP. Images are fed to the model in one of three ways — as base64-encoded bytes in the request body, as an external URL the model fetches, or as a file_id reference through the new Files API. The per-request limits published in the Vision guide allow up to 600 images per call, with a maximum resolution of 8,192 pixels per side for requests containing fewer than 15 images, dropping to 4,096 pixels when 15 or more images are present.

The model supports thinking mode and has it enabled by default, inheriting V4-Flash's chain-of-thought reasoning behavior. It is accessible through both DeepSeek's OpenAI-compatible endpoint and its Anthropic-format endpoint, meaning teams already using either format can add vision by changing the model identifier with no structural changes to their API integration.

The Benchmark Table the Coverage Got Wrong

The most cited fact from the launch is that V4-Flash-Vision-Exp performs "close to Opus 4.8" on multimodal agent benchmarks. That phrase is DeepSeek's own, and it is accurate for the multimodal-specific rows in the comparison table. The problem is that the table has eleven rows, not four, and the model's standing across all eleven is considerably less flattering than the headline comparison suggests.

DeepSeek's release note embedded the benchmark data as an image file rather than text, making it inaccessible to many text-scraping tools and therefore underanalyzed in early coverage. The full table, self-reported by DeepSeek and comparing V4-Flash-Vision-Exp, the text-only V4-Flash-0731, and Anthropic's Opus 4.8, breaks into two sections.

On the seven text-based agent benchmarks, V4-Flash-Vision-Exp outperforms its text-only baseline on six and trails on one (Cybergym). Against Opus 4.8, however, the vision model leads only on DeepSWE — a software engineering benchmark — by 1.3 points (59.3 versus 58.0), and trails on the remaining six. The gap on NL2Repo, which evaluates natural-language-to-repository code generation, runs to 12.0 points (57.7 versus 69.7). On DSBench-Hard, the gap is 8.1 points.

The four multimodal-agent benchmarks are where the "close to Opus 4.8" characterization has the strongest foundation. The vision model leads Opus 4.8 on Agents' Last Exam (27.3 versus 25.7) and ZeroBench (35.0 versus 34.0), and trails narrowly on ApexBench (36.5 versus 39.4) and Chartography (64.3 versus 65.0). These are the benchmarks built to test exactly the capability the model just added — the ability to act on visual content — so leading or running close on four of four multimodal tasks is a meaningful result. But it does not translate into general parity with a model that holds an eight-to-twelve point advantage on the harder text-agent evaluations.

Critical context: all eleven scores are self-reported by DeepSeek. The text-agent rows were evaluated using DeepSeek's own Harness framework in Minimal Mode, with specific hyperparameter settings (temperature 1.0, top_p 0.95). The source of the Opus 4.8 comparison scores is not disclosed. No independent evaluator has published a reproduction of any score in the table. The Terminal Bench 2.1 leaderboard showed no third-party entry for this model as of publication. Benchmark precedent from the broader V4 family is not encouraging: as documented by analysts tracking the V4-Pro launch, V4-Pro's preview build registered 80.6 percent on DeepSeek's own SWE-bench Verified evaluation but only 8 percent on the independently run DeepSWE benchmark, a gap attributed primarily to the difference in verifier false-positive rates between the two tests.

Two Clocks, One Rate Card: How Image Billing Actually Works

Pricing for V4-Flash-Vision-Exp is identical to text-only V4-Flash on every line item, with one significant complication most coverage has ignored: DeepSeek charges two different rates depending on the time of day — but only on weekdays.

The off-peak rate — in effect on weekends and during the non-peak hours of weekdays — is $0.22 per million input tokens on a cache miss, $0.007 per million on a cache hit, and $0.66 per million output tokens. During two peak windows on weekdays (01:00 to 04:00 UTC and 06:00 to 10:00 UTC, Monday through Friday), every line item doubles exactly: $0.44 input, $0.014 cache hit, $1.32 output. Weekends are entirely off-peak under DeepSeek's published rate structure. A pipeline running during London business hours (09:00 to 17:00 UTC on weekdays) operates almost entirely off-peak. A pipeline running overnight US East Coast time on weekdays (01:00 to 05:00 ET is 05:00 to 09:00 UTC) catches most of the second peak window and pays double rates for that period.

Image tokens are billed at whatever input rate applies at the moment of the call — peak or off-peak, cache miss or cache hit — with no vision-specific surcharge. Using the 384-token ceiling and the off-peak cache-miss rate, a single image costs approximately $0.0000845. One thousand images cost roughly $0.085. At scale, a pipeline processing one million images costs approximately $84.50 in image-token input alone at off-peak rates — a number dwarfed by the output tokens the model produces in response to those images, which at even a modest 200 output tokens per image would cost $132 at off-peak rates for the same million-image batch. The practical reading is that image tokens are economically negligible relative to the output they generate; the clock and the thinking-mode default are the more consequential cost factors for any production deployment.

The Files API, which shipped on the same day as the vision model, allows teams to upload an image once and reference it by file_id across subsequent requests. DeepSeek states it is free. The practical benefit is request size: without the Files API, the total image payload per request is capped at 64 MiB; when images are referenced via file_id, the total image size limit rises to 200 MiB per request. For teams processing high-resolution screenshots in agent loops where the same source images recur across turns, the Files API eliminates repeated transfers without requiring re-encoding. Whether repeated file_id references also trigger DeepSeek's context-cache discount on the token billing is not documented; the two mechanisms are separate systems and cannot be assumed equivalent.

What the Model Cannot Do

Several capabilities absent from V4-Flash-Vision-Exp are worth naming explicitly, both because they appear in competing multimodal models and because early coverage implied a broader capability set than the documentation supports.

There is no video input, no native PDF or document ingestion, no OCR mode, and no spatial grounding or bounding-box output. Input is text and still images; output is text. Teams that need page-level document understanding must rasterize PDF pages into image files and send them within the 600-image, 64-MiB payload limits — which is technically possible but requires the document processing layer to be built and maintained by the developer.

The model has no open weights. DeepSeek's text-only V4-Flash checkpoint is available on Hugging Face under an MIT license and can be self-hosted. The vision model has no corresponding repository, no license, and no announced path to open weight release. Teams choosing DeepSeek for its self-hosting option — typically to address data-sovereignty concerns — cannot replicate that option for the vision capability. For the duration of the experimental period, the model is API-only.

The "Exp" suffix is worth taking seriously. DeepSeek has used the designation before, including on DeepSeek-V3.2-Exp. No formal definition of what Exp commits the company to — in terms of stability, evaluation timeline, or deprecation policy — has been published. The practical implication is that the model identifier deepseek-v4-flash-vision-exp should be treated as configuration rather than a permanent string. When the Exp period ends, DeepSeek typically moves the capability to a stable identifier, and teams with a hard-coded model string may discover the switch through an API error.

Competitive Context: Closing the Vision Gap in Chinese AI Agent Infrastructure

V4-Flash-Vision-Exp is not the first multimodal model from a major Chinese AI lab, and it is not the cheapest available vision model. Gemini 3.5 Flash has offered multimodal capabilities at comparable price points. Qwen-VL from Alibaba offers open weights with vision support. What the DeepSeek release represents is the closure of a specific gap that was costly for a specific developer community: teams that had committed to DeepSeek's agent ecosystem, built on V4-Flash and the Harness framework, and were absorbing the overhead of a two-model pipeline to handle images.

For those teams, the architecture simplification is the primary value, not the benchmark comparison. A single model endpoint that handles both text and images with a uniform pricing model and native Harness integration removes a decision point, reduces latency from the inter-model handoff, and eliminates a category of failure mode in which the vision extraction step returns something the text model misinterprets.

For teams without a prior DeepSeek dependency, the competitive calculus is more open. Gemini 3.5 Flash offers multimodal with a more established production track record and a tile-based image pricing model that scales with resolution — in some workloads lower than DeepSeek's 384-token ceiling, in others higher, depending on image complexity and the tile resolution selected. The Opus 4.8 comparison in DeepSeek's benchmark table is the wrong comparison for most cost-sensitive multimodal agent developers; the more relevant question is how V4-Flash-Vision-Exp performs against Gemini Flash in independent evaluations on actual agent tasks — a test that will require third-party reproducers to run.

The data-sovereignty dimension remains unchanged and extends to the vision context in a meaningful way. DeepSeek is incorporated in Hangzhou, China and owned by High-Flyer, a Chinese hedge fund. China's National Intelligence Law requires organizations to cooperate with government intelligence requests; the Cybersecurity Law and Data Security Law impose additional data access and transfer obligations. When the hosted API processes images — screenshots of business applications, document scans, interface states — the scope of potential data exposure expands beyond text content. Enterprise teams routing sensitive visual content through the API inherit the same obligations they would for text, applied now to image inputs that may capture application internals, financial documents, or proprietary UI designs. Self-hosting is not currently an option for the vision model.

What Comes Next for DeepSeek's Multimodal Stack

The most important development to watch is independent benchmark replication. DeepSeek's eleven-benchmark self-reported table is internally more honest than the coverage it received — a lab that wanted to claim full parity with Opus 4.8 could have published a narrower table — but the scores tell a more complex story than the headline comparison, and DeepSeek's prior benchmark discrepancies between vendor-controlled and independently-run evaluations make third-party verification a prerequisite for any serious production-planning decision.

A forthcoming API price increase announced by DeepSeek — with no disclosed timeline, magnitude, or scope — is the second open variable. Every cost figure in this article reflects rates that DeepSeek has explicitly characterized as temporary. Teams modeling a cost case for adopting V4-Flash-Vision-Exp should treat current pricing as a snapshot rather than a steady-state, and should model the scenario where peak-hour rates become the new off-peak rates.

The third question is whether DeepSeek releases the vision checkpoint as open weights. If it does, teams currently blocked by data-sovereignty requirements from using the hosted API would gain access to the vision capability on self-hosted infrastructure. If it does not, the vision model will remain a proprietary-only capability in a product line otherwise known for open weight releases — a structural difference from V4-Flash that the developer community will eventually need to incorporate into its platform decisions.

What the August 21 release has already established is that DeepSeek's agent model can now see. Whether that vision is good enough, at what price, under what data conditions, and for which workloads — those are the questions the coming weeks of independent evaluation will begin to answer.

How Mixture-of-Experts Architecture Makes Low-Cost Vision Possible

The economic logic behind V4-Flash-Vision-Exp's pricing sits in the MoE architecture that V4-Flash inherited from DeepSeek's earlier models. A conventional dense language model — one in which every parameter participates in every forward pass — scales inference cost linearly with parameter count. A 284 billion parameter dense model would require moving 284 billion parameter values through hardware accelerators for every token the model processes. That is prohibitive at consumer pricing.

A sparse MoE model decouples total parameter count from per-token compute. V4-Flash's 284 billion parameters are organized into specialized "expert" subnetworks. When the model processes a token, a learned routing mechanism selects a small subset of experts — roughly 13 billion parameters' worth — and routes the token through only those experts. The remaining parameters are dormant for that particular token. Because different tokens route to different experts, the full 284 billion parameter knowledge base is exercised across a diverse workload, but any single inference step touches only a fraction of the total parameters. DeepSeek first used this design in V2, developed it further in V3, and retained it as the core architectural principle in V4.

Two attention mechanisms introduced in the V4 architecture add further compute savings specifically at long context lengths: Compressed Sparse Attention, which groups nearby tokens and retrieves only statistically relevant long-distance context, and Heavily Compressed Attention, which represents older context as dense summary vectors. These mechanisms reduce per-token computation at very long sequences, making the million-token context window economically viable rather than merely technically possible.

Adding a vision encoder to this foundation does not fundamentally change the text inference economics. DeepSeek has not described its vision pathway, but the standard approach for multimodal models built on transformer backbones is to project image tokens into the same embedding space the language model uses for text, then process the combined sequence through the model's existing layers. The 384-token ceiling DeepSeek applies per image is consistent with this approach: it caps the number of visual tokens that enter the sequence, bounding the compute cost per image while preserving the image's informational content at a resolution appropriate for most business documents and screenshots.

The Agent Framework That Makes the Vision Model Immediately Deployable

DeepSeek Harness is the agent runtime that DeepSeek has developed alongside its V4 model family. Version 0.1 launched alongside the V4-Pro GA release on August 13, 2026; v0.1.1 shipped on the same day as the vision model and adds out-of-the-box support for deepseek-v4-flash-vision-exp.

The significance of same-day Harness support is that the developer friction for adding image processing to an existing DeepSeek-based agent is minimal. Developers already using Harness with V4-Flash change the model identifier in their configuration and add image inputs to their prompts. The framework handles the rest: tool integration, multi-turn memory management, and the task-loop infrastructure that allows agents to operate across extended sequences of actions. For the agentic use cases where DeepSeek V4-Flash built its developer following — coding agents, terminal-based task execution, browser automation — the vision extension means the agent can now read the screen state it is acting on, rather than inferring it from text output alone.

The types of workflows that become architecturally simpler include browser agents that must read rendered page states, UI test automation tools that verify visual outputs, document-processing pipelines that handle scanned forms or screenshots of structured data, and coding agents that need to interpret error screenshots or UI mockups. These are not hypothetical applications — they represent the actual workflows for which the two-model workaround was most painful, and they correspond directly to the multimodal-agent benchmarks (ApexBench, Agents' Last Exam) where DeepSeek's self-reported data shows the strongest improvement over the text-only baseline.

Comparing Vision Pricing Across the Tier

V4-Flash-Vision-Exp is not the only option for developers looking for low-cost multimodal inference, and a direct comparison is necessary to evaluate DeepSeek's pricing claim accurately.

Google's Gemini 3.5 Flash is the most established competitor at a similar pricing tier. Gemini 3.5 Flash charges image input separately from text input, applying a per-image tile fee that depends on resolution and tile count. The critical difference from DeepSeek's approach is that Gemini's per-image cost scales with resolution under its tile-based model, while DeepSeek's 384-token ceiling is flat regardless of resolution above a threshold. For batch workloads processing high-resolution architectural diagrams or large documents, DeepSeek's flat ceiling may be more predictable. For workloads processing many small, simple images, Gemini's tile pricing may be lower per image. A direct cost comparison requires measuring against actual workload images rather than theoretical ceilings.

Other low-cost hosted vision APIs are available from OpenAI and other providers, with output pricing that may favor DeepSeek's off-peak rates in some deployment windows. Alibaba's Qwen-VL models offer open-weight multimodal options that can be self-hosted, making them functionally distinct from the hosted-API comparison: a team that can run Qwen-VL on its own GPU infrastructure faces marginal inference costs rather than per-token API costs at scale. DeepSeek's vision model, with no open weights, does not compete in that category.

The honest competitive summary: V4-Flash-Vision-Exp is priced competitively within the hosted multimodal API market, its pricing mechanics are simpler than some alternatives (flat 384-token image ceiling versus tile-based or resolution-scaled charging), and its pricing advantage over frontier multimodal models is substantial — but the absence of open weights means the total cost advantage versus self-hosted options depends entirely on workload volume and infrastructure economics.

Independent Verification and What Developers Should Do Before Committing

The absence of independent benchmark reproduction is not a minor editorial footnote — it is the central unresolved question for any developer considering a production commitment to V4-Flash-Vision-Exp. DeepSeek's model family has a documented track record of discrepancy between vendor-controlled benchmark scores and independent evaluation scores. As analysts noted at the V4-Pro GA launch, the V4-Pro preview registered 80.6 percent on DeepSeek's own SWE-bench Verified evaluation, which uses a verifier with an 8.5 percent false-positive rate; on the independently run DeepSWE benchmark, which uses a 0.3 percent false-positive verifier from yage.ai, the same model scored 8 percent.

The 0813 post-training update moved V4-Pro's DeepSWE score to 62.7 under the tighter benchmark — suggesting genuine capability improvement — but that score is itself still vendor-reported. The vision model's multimodal benchmark scores have not yet been subjected to even this level of external scrutiny.

The practical implication for developers is to treat the benchmark table as directional rather than definitive, and to run the model against their own representative workloads before committing to integration at scale. ApexBench and Agents' Last Exam, which form the core of DeepSeek's multimodal comparison, are designed to test agent behavior on complex multi-step visual tasks; results on those benchmarks may not transfer to the specific screenshot types, document formats, or UI patterns that a given team's workload produces. A team building a data extraction pipeline for scanned invoices will learn more from feeding a representative sample of its own invoices to the model than from the Chartography benchmark score. The cost of a preliminary evaluation is negligible — at off-peak rates, processing a thousand test images costs less than ten cents in image tokens — and the information value is substantially higher than any score in the self-reported table.