
News
DeepSeek-V4-Flash-Vision-Exp Release: Multimodal API Now Live | DeepSeek API Docs
2026.08.28
·Web·by Mineru#Agent#API#LLM#Multimodal AI#Vision Model
Key Points
- 1DeepSeek has launched the experimental `deepseek-v4-flash-vision-exp` model, which integrates advanced multimodal capabilities while maintaining parity with DeepSeek-V4-Flash’s text and reasoning performance.
- 2The model supports mixed text and image inputs via base64, URLs, or the new Files API, with images tokenized at a fixed rate of 384 tokens for efficient billing.
- 3Designed for agentic workflows, the release includes updated framework support and a free Files API to allow for image reuse across multiple requests, significantly reducing bandwidth requirements.
The release of DeepSeek-V4-Flash-Vision-Exp represents a significant expansion of the DeepSeek-V4 ecosystem, introducing a multimodal architecture that integrates advanced visual understanding with the core agentic, reasoning, and world-knowledge capabilities of the V4-Flash model.
Core Architecture and Performance
The model is engineered to provide feature parity with DeepSeek-V4-Flash regarding text-based tasks while achieving a substantial performance boost in multimodal agentic benchmarks. By augmenting the existing text-heavy V4-Flash framework with visual processing capabilities, the model nears the performance levels of industry-standard benchmarks such as Opus-4.8. This integration is designed to facilitate complex agentic workflows that require the synthesis of visual input with tool-use frameworks.Multimodal Methodology and Technical Specifications
- Tokenization and Billing: The model employs a specialized visual tokenization scheme for billing purposes, capping the cost at a maximum of 384 tokens per image. The pricing structure is harmonized with the standard V4-Flash model to maintain economic efficiency.
- Input Modality: The architecture supports mixed-input streams, allowing for the concurrent processing of text and visual data. The API is compatible with multiple ingestion methods:
- Base64 encoding: Direct embedding of image data.
- External URLs: Fetching assets from remote storage.
- Files API: Utilizing persistent storage for efficient data management.
- Integration Framework: The model is fully integrated with *DeepSeek Harness 0.1.1*, enabling seamless deployment within existing agentic frameworks. It supports standard API protocols, including Chat Completions, Messages, and Responses.
Files API and Optimization
To mitigate bandwidth constraints and redundant processing, the Files API has been introduced as an auxiliary utility. This system operates on a persistent storage mechanism where users can upload an image once and reference it via a uniquefile_id. This implementation allows for:
- Bandwidth Efficiency: Reducing the latency and data overhead associated with repeated transmission of visual inputs.
- Resource Management: Enabling multiple consecutive requests to reference the same visual context without re-uploading, which is critical for long-running agentic processes.
- Cost Optimization: The Files API is currently provided at no additional cost, incentivizing the reuse of visual context within multi-turn interactions.