DeepSeek has unveiled an experimental multimodal model designed to bring image understanding and visual interaction capabilities to its V4 Flash family, intensifying competition with Anthropic’s Claude Opus 4.8. Called DeepSeek-V4-Flash-Vision-Exp, the model can process images and screenshots in addition to handling the text-based reasoning, coding and agent capabilities of the existing V4 Flash model.
The Chinese AI startup says the experimental model approaches or surpasses Anthropic’s Opus 4.8 on several visual and agent benchmarks. DeepSeek’s published results show the model beating Opus 4.8 on two visual benchmarks, while other tests show it narrowing the gap with Anthropic’s frontier system. The release adds another dimension to the intensifying AI competition, where model capability is increasingly being measured alongside inference costs and efficiency.
DeepSeek V4 Flash Vision Exp Adds Multimodal Capabilities
DeepSeek-V4-Flash-Vision-Exp is an experimental version of the company’s V4 Flash model.
The key upgrade is the addition of vision capabilities. While the original V4 Flash was primarily a text-based model, the new system can interpret visual inputs such as images and screenshots.
DeepSeek said the experimental model retains the text capabilities of V4 Flash, including reasoning, agentic tasks and world knowledge.
DeepSeek’s New Model At A Glance
| Metric | DeepSeek-V4-Flash-Vision-Exp |
|---|---|
| Model family | DeepSeek V4 Flash |
| Model type | Experimental multimodal model |
| Text capabilities | Yes |
| Image understanding | Yes |
| Screenshot analysis | Yes |
| Agentic tasks | Yes |
| Reasoning | Yes |
| Coding | Yes |
| Availability | DeepSeek developer/API platform |
| Status | Experimental |
The model is currently available through DeepSeek’s paid developer platform, according to SiliconANGLE. The company could potentially make a broader or free version available later.
DeepSeek Says It Beats Opus 4.8 On Some Visual Tests
The most notable aspect of the launch is DeepSeek’s comparison with Anthropic’s Claude Opus 4.8.
DeepSeek tested its experimental model against several leading systems across visual and agentic benchmarks. Its published results show the new model outperforming Opus 4.8 on two visual benchmarks, including ALE and ZeroBench.
DeepSeek Vs Opus 4.8
| Benchmark | DeepSeek-V4-Flash-Vision-Exp | Claude Opus 4.8 | Result |
|---|---|---|---|
| ALE | Higher | Lower | DeepSeek ahead |
| ZeroBench | Higher | Lower | DeepSeek ahead |
| ApexBench | 36.5 | 39.4 | Opus 4.8 ahead |
| Other text/agent benchmarks | Mixed | Mixed | Varies |
The benchmark results are based on DeepSeek’s own evaluation and should not automatically be interpreted as evidence that the model is universally more capable than Opus 4.8.
ApexBench Shows A Narrow Gap
One benchmark cited in coverage of the launch is ApexBench.
DeepSeek-V4-Flash-Vision-Exp scored 36.5 at Pass@1, compared with 39.4 for Claude Opus 4.8. That places DeepSeek relatively close to Anthropic’s model on the particular evaluation.
ApexBench Comparison
| Model | ApexBench Pass@1 |
|---|---|
| DeepSeek-V4-Flash-Vision-Exp | 36.5 |
| Claude Opus 4.8 | 39.4 |
| Difference | 2.9 points |
The result suggests that DeepSeek’s new model is competitive on at least some demanding multimodal workloads, although benchmark scores should be evaluated alongside independent testing, real-world performance and cost.
ALE Tests Complex Visual Agent Tasks
ALE is particularly relevant because it is designed around tasks requiring AI systems to interact with applications, write code and interpret media files.
The benchmark contains more than 1,000 multi-step tasks, according to SiliconANGLE.
That makes it different from a simple image-recognition test.
A model needs to understand visual information and use that information as part of a broader sequence of actions.
What Visual Agents Need To Do
| Capability | Importance |
|---|---|
| Read screenshots | Understand application state |
| Interpret images | Extract visual information |
| Reason over multiple steps | Complete complex tasks |
| Write code | Automate or modify workflows |
| Use tools | Interact with applications |
| Maintain context | Track task progress |
DeepSeek’s performance on such tests is significant because AI companies increasingly see visual agents as a major application area.
ZeroBench Focuses On Difficult Image Analysis
DeepSeek also compared the model on ZeroBench, which contains 100 challenging image-analysis tasks designed for frontier AI systems.
DeepSeek-V4-Flash-Vision-Exp reportedly outperformed Opus 4.8 on this evaluation.
The benchmark is intended to test difficult visual reasoning rather than basic object recognition.
That could make the result relevant for applications involving documents, diagrams, screenshots and other complex visual inputs.
The Model Builds On DeepSeek V4 Flash
The experimental vision model is derived from DeepSeek V4 Flash.
DeepSeek released V4 Flash earlier in 2026 as part of its V4 model family. The underlying model uses a mixture-of-experts architecture, according to SiliconANGLE’s analysis of DeepSeek’s published V4 Flash information.
V4 Flash Architecture
| Specification | Reported Detail |
|---|---|
| Total parameters | 284 billion |
| Parameters per expert network | 13 billion |
| Architecture | Mixture of experts |
| Training data | 32 trillion tokens |
| Context processing | KV-cache compression |
| Context supported | Up to 1 million tokens |
A mixture-of-experts architecture allows a model to activate only selected portions of its neural network for a particular request rather than running the entire parameter set for every token.
This can reduce the computing resources required for inference.
DeepSeek Focuses On Efficiency
Efficiency has become one of DeepSeek’s defining characteristics.
The company has repeatedly emphasized achieving strong AI performance without requiring the same level of computational resources associated with some frontier models.
V4 Flash uses two techniques, HCA and CSA, to compress its KV cache. SiliconANGLE reported that DeepSeek says the methods reduce computing requirements for processing prompts containing one million tokens by 73%.
Efficiency Features
| Technology | Purpose |
|---|---|
| Mixture of experts | Activates selected neural networks |
| HCA | KV-cache compression |
| CSA | KV-cache compression |
| Muon | Speeds training optimization |
| 1M-token context | Supports large inputs |
These techniques are important because AI companies are increasingly competing not just on benchmark performance but on the amount of useful work that can be completed per dollar of compute.
DeepSeek’s Vision Model Targets AI Agents
The new model’s visual capabilities are particularly relevant to AI agents.
An agent that can understand only text has limited visibility into graphical interfaces. A vision-capable agent can potentially read a screen, identify buttons and interface elements, interpret images and then take actions based on what it sees.
This opens up applications across software development, office automation, research and customer support.
Potential Applications
- Computer-use agents
- Software development
- Screenshot analysis
- Document processing
- Visual research
- Browser automation
- UI testing
- Data extraction
- Image-based reasoning
The ALE benchmark’s emphasis on application interaction and coding reflects this emerging use case.
DeepSeek Is Also Competing On Price
The launch comes during an increasingly aggressive AI pricing race.
DeepSeek’s V4 Flash has been positioned as a low-cost model, while other AI providers have increasingly reduced prices or introduced cheaper models to compete for developer workloads.
Coverage of the new vision model reported that processing 1,000 images can cost approximately 1.15 yuan through the DeepSeek API, or around $0.17, although actual costs can vary based on token usage and API pricing conditions.
Reported Image Processing Economics
| Metric | Reported Figure |
|---|---|
| Images | 1,000 |
| Approximate processing cost | 1.15 yuan |
| Approximate US-dollar equivalent | $0.17 |
| Model | V4-Flash-Vision-Exp |
| Availability | API |
If such economics are maintained at scale, DeepSeek could put additional pressure on developers of premium multimodal models.
Anthropic’s Opus 4.8 Remains A Major Benchmark
Anthropic’s Claude Opus 4.8 is positioned as a high-end model with strong coding and agentic capabilities.
Anthropic says Opus 4.8 delivers improvements across coding and agentic tasks and has undergone extensive alignment testing.
DeepSeek’s decision to use Opus 4.8 as a comparison point demonstrates how quickly Chinese AI models are moving toward direct competition with leading Western systems.
DeepSeek Vs Anthropic
| Area | DeepSeek V4 Flash Vision Exp | Claude Opus 4.8 |
|---|---|---|
| Text reasoning | Yes | Yes |
| Image understanding | Yes | Yes |
| Agentic tasks | Yes | Yes |
| Coding | Yes | Yes |
| Visual benchmarks | Competitive | Frontier-level |
| Pricing strategy | Low-cost focus | Premium frontier positioning |
| Model status | Experimental | Production frontier model |
The comparison is therefore about more than benchmark scores. Developers are increasingly evaluating models based on the combination of capability, reliability, speed and cost.
DeepSeek’s Release Comes After V4 Pro
The new vision model arrives shortly after DeepSeek released the general-availability version of V4 Pro.
DeepSeek V4 Pro-0813 was released in August and has also been compared with leading frontier systems on agent and coding benchmarks.
This suggests DeepSeek is expanding the V4 family quickly rather than relying on a single general-purpose model.
DeepSeek V4 Family
| Model | Focus |
|---|---|
| V4 Flash | Fast, lower-cost general workloads |
| V4 Pro | Higher-end general and agentic workloads |
| V4 Flash Vision Exp | Experimental multimodal and visual agents |
A future vision-optimized version of V4 Pro could potentially target even more demanding multimodal workloads, although DeepSeek has not confirmed such a product.
Why Multimodal AI Matters
The AI industry is increasingly moving beyond text-only interaction.
Humans communicate and work through a mixture of text, images, screens, documents, video and other formats. AI systems that can process these different types of information can potentially operate more effectively in real-world environments.
For example, a software agent could inspect a screenshot of an error, read the relevant code and then modify the application.
Similarly, an enterprise agent could interpret a scanned document, extract information and enter it into another application.
The Competitive Gap Is Becoming Harder To Define
DeepSeek’s latest release highlights how difficult it is becoming to classify AI models simply as “leading” or “behind.”
A model may outperform another system on one benchmark while trailing it on another.
DeepSeek-V4-Flash-Vision-Exp provides a good example: it reportedly beats Opus 4.8 on some visual evaluations while remaining behind it on others, such as ApexBench.
What Developers Need To Consider
| Evaluation Factor | Why It Matters |
|---|---|
| Benchmark performance | Measures specific capabilities |
| Real-world reliability | Determines production usefulness |
| API price | Affects operating costs |
| Latency | Affects user experience |
| Context window | Enables larger tasks |
| Tool use | Important for agents |
| Multimodal support | Enables visual workflows |
| Safety | Important for enterprise adoption |
This means benchmark leadership on a single test is unlikely to determine which model wins commercially.
DeepSeek’s Open-Model Strategy Could Matter
DeepSeek has previously open-sourced several models, helping developers and researchers access its technology directly.
The latest experimental vision model is currently being offered through the developer platform rather than as a broadly available open-weight release, according to SiliconANGLE.
Whether DeepSeek eventually releases the vision model’s weights could significantly affect its adoption among developers who prefer running AI models themselves.
The Bigger Picture
DeepSeek’s V4-Flash-Vision-Exp represents another step in the rapid convergence between Chinese and Western frontier AI capabilities. The experimental model adds image and screenshot understanding to DeepSeek’s V4 Flash platform and, according to DeepSeek’s own benchmark results, outperforms Anthropic’s Opus 4.8 on some visual evaluations while remaining competitive on other demanding tests.
The bigger significance is the combination of capability and economics. DeepSeek is attempting to deliver increasingly sophisticated multimodal and agentic functionality while maintaining a low-cost model strategy. If developers find the model reliable in real-world applications, it could put additional pressure on premium AI providers to improve both performance and pricing.
Looking Ahead
The next important test for DeepSeek-V4-Flash-Vision-Exp will be independent evaluations and real-world developer adoption. DeepSeek’s published benchmark results show strong performance in several visual tasks, but developers will ultimately judge the model based on reliability, latency, tool use, image understanding and total cost of completing real tasks. Its experimental status also means that performance and availability could change as DeepSeek continues to develop the system.
For the wider AI industry, the launch reinforces the growing importance of multimodal agents. The competition is no longer limited to which model can answer the most difficult text questions; companies are increasingly building systems that can understand screens, operate software and complete multi-step tasks. DeepSeek’s latest model shows that Chinese AI companies are continuing to push aggressively into this frontier, potentially accelerating both the capability race and the price competition across the global AI market
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.
