DeepSeek-V4-Flash-Vision-Exp is an experimental multimodal model developed as an extension of the DeepSeek-V4-Flash architecture. The model is engineered by integrating specialized visual modules into the existing flash-based foundation, followed by a continued training phase specifically designed to optimize visual perception and reasoning capabilities.

Core Methodology and Architecture

The model’s architecture builds upon the DeepSeek-V4-Flash framework, incorporating several advanced structural components:

1. Multimodal Integration: The architecture utilizes a dedicated vision encoder and aligner to map visual input into the model’s latent space. This allows the model to process multimodal data while maintaining the text-only performance characteristics of the original V4-Flash model.
2. DFlash Attention & MoE: The model leverages DFlash attention mechanisms combined with a Mixture-of-Experts (MoE) architecture. This enables high-efficiency inference by activating only a subset of parameters per token, balancing computational load with performance.
3. Hyper-Connections & DSpark: The inference path includes advanced structural features known as Hyper-Connections and the DSpark forward path. DSpark is particularly utilized for speculative decoding during the generation phase.
4. Speculative Decoding (DSpark): The model supports a specific speculative implementation called DSPARK. In this regime, the target and draft models originate from the same checkpoint, utilizing a probabilistic draft sampling method with adaptive verification to accelerate generation throughput.
5. Precision and Storage: The model architecture handles diverse tensor types, including $BF16$, $F32$, $F8\_E4M3$, $I8$, and $I64$. The total parameter count is approximately $305B$, managed via sharded Safetensors files to optimize loading and inference distribution.

Technical Capabilities

DeepSeek-V4-Flash-Vision-Exp demonstrates significant performance enhancements in multimodal agent tasks, as evidenced by its scoring on benchmarks such as ApexBench (36.5) and Agents' Last Exam (27.3).

The framework is highly optimized for deployment environments:

• Prompt Encoding: The model supports both OpenAI-style JSON content blocks and compact text notation, defined as $\langle image \rangle path \langle /image \rangle$.
• Inference Optimization: By employing $FP8$ quantization for the Key-Value (KV) cache and implementing hardware-aware configurations (e.g., $--block-size \ 256$), the model maintains high tokens-per-second performance.
• Agent Framework: The model is compatible with the DeepSeek Harness agent framework, supporting complex tool-calling and reasoning chains through specialized parsers (--tool-call-parser and --reasoning-parser set to deepseek_v4).

In summary, the model represents a specialized refinement of the DeepSeek-V4-Flash line, prioritizing the fusion of high-performance text-based reasoning with visual interpretation, while providing a modular repository layout to facilitate flexible deployment across varied compute infrastructures using tools like vLLM and SGLang.