DeepSeek Releases V4 Flash Vision Exp, an Experimental Multimodal MoE Model with 1M Context
DeepSeek has released V4 Flash Vision Exp, an experimental vision-enabled variant of DeepSeek V4 Flash 0731 that adds image understanding while matching the base model's text performance. The sparse mixture-of-experts model uses 13B active parameters out of 284B total and supports a 1M token context window.
DeepSeek V4 Flash Vision Exp at a glance
How DeepSeek V4 Flash Vision Exp ranks now
| Benchmark | Score | Rank | Source |
|---|---|---|---|
| GPQA | 91.3% | 34 of 165 | Artificial Analysis, Oct 2026 ↗ |
Rank is against every model in our database with a sourced score on that benchmark, each counted once with its best score.
DeepSeek Adds Vision to V4 Flash Line
DeepSeek has released DeepSeek V4 Flash Vision Exp, an experimental vision-enabled version of its DeepSeek V4 Flash 0731 model. The new model adds image understanding capabilities while, according to DeepSeek, matching the base text model's performance on agents, reasoning, and world knowledge tasks.
Architecture and Specs
V4 Flash Vision Exp is a sparse mixture-of-experts (MoE) model with 13B active parameters out of 284B total parameters. This active/total ratio is consistent with DeepSeek's established approach of using large parameter pools while keeping inference costs low through sparse activation.
The model supports a 1M token context window, positioning it for long-document and multi-image workflows. It is listed as available via OpenRouter, with a release date of August 21, 2026.
Pricing
OpenRouter lists pricing at:
- Input: $0.22 per 1M tokens
- Output: $0.66 per 1M tokens
- Cache read: $0.007 per 1M tokens
No separate benchmark scores have been published for this release. DeepSeek has not disclosed formal evaluation results comparing V4 Flash Vision Exp against other multimodal models on standard benchmarks like MMMU or DocVQA.
Intended Use Cases
DeepSeek positions the model for:
- Document and chart understanding
- Visual question answering
- Multimodal agent workflows that interleave text and images
The "Exp" designation in the model name signals this is an experimental release, suggesting DeepSeek is testing the vision capability ahead of a potential stable version. The model builds directly on DeepSeek V4 Flash 0731, extending it rather than representing a ground-up architecture change.
What this means
This release extends DeepSeek's Flash line into multimodal territory without, according to the company, sacrificing the text-only model's reasoning and agentic capabilities — a claim that has not been independently verified through published benchmarks. The 13B active-parameter design keeps inference costs comparable to DeepSeek's other efficient models, and the sub-$1 per-million-token pricing on both input and output continues the aggressive cost positioning that has characterized DeepSeek's strategy against Western frontier labs.
The 1M token context window is notable for a vision-capable model at this price point, potentially enabling use cases like analyzing lengthy multi-page documents with embedded charts and images in a single pass. However, the experimental tag means enterprises should expect further iteration before this becomes a production-recommended model. Real-world performance on multimodal benchmarks will determine whether it can compete with established vision models from OpenAI, Google, and Anthropic, none of which have published comparable figures yet for this specific release.
Related Articles
Thinking Machines Lab releases Inkling 2 Mini: 26B-parameter looped MoE with 3B active and 1M-token context
Thinking Machines Lab has released Inkling 2 Mini, a multimodal Mixture-of-Experts model with about 26B physical total parameters and 3B active per token. It repeats its middle layers three times, supports a 1M-token context, and ships with a DSpark drafter. Hugging Face reports it leads 32 of 48 benchmarks against four comparable models.
Tencent releases Hy4 flagship weights: 770B-parameter MoE, 1,048,576-token context, conditional license
Tencent released downloadable weights for its flagship Hy4 on November 20, with a System Card. The multimodal MoE has a 770-billion-parameter language backbone, 49 billion active parameters per text token and a 1,048,576-token context window, under a license with commercial conditions for model-as-a-service operators.
Tencent releases Hy4 flash: 220B MoE, 12B active parameters, 1M-token context
Tencent has released Hy4 flash, a multimodal Mixture-of-Experts model with 220B total parameters, 12B active per token, and a 1,048,576-token context window. Tencent calls it its most cost-effective Hy model to date and claims a per-task cost of 1/20 that of Hy4 on its internal WorkIndex benchmark.
DeepSeek removes separate fees for model-triggered search and web fetch from Nov 18
DeepSeek removed separate charges for server-side search and web fetch triggered during model calls, effective at noon on November 18, 2026, Beijing time. Token usage is still billed, and the standalone Search and Fetch APIs remain at 5 yuan per 1,000 successful calls.
Comments
Loading...