DeepSeek-V4-Flash-0731-Vision

An experimental vision-language initialization combining DeepSeek-V4-Flash-0731 with Kimi-K3's vision encoder through a cross-attention adapter.

This release is intended as a starting point for further multimodal training and research rather than a finalized model.

Model details

Component	Description
Language model	DeepSeek-V4-Flash-0731
Vision encoder	Kimi-K3
Vision integration	Cross-attention adapter

The repository contains the language-model weights, vision encoder, and cross-attention adapter.

By design, text-only requests bypass the vision encoder and cross-attention adapter and use the frozen DeepSeek language-model path unchanged.

Evaluation

This initialization scores 54.56% on MMMU validation with high reasoning effort.

For broad context, here are two published reference points:

Model	MMMU validation
Pixtral 12B	52.5%
DeepSeek-V4-Flash-0731-Vision	54.56%
Qwen2.5-VL-7B-Instruct	58.6%

Scores use each model's reported setup and are shown only for context.

Limitations

This is an initialization checkpoint rather than a finalized vision-language model. Fine-grained spatial grounding, very small text, and OCR may be inconsistent. Results can also vary with image preprocessing, prompting, and decoding settings.

Additional multimodal training may improve these areas.

License

MIT.

This repository includes components from DeepSeek-V4-Flash-0731 and Kimi-K3. Redistributed upstream weights remain subject to their original licenses, which are not overridden by this license.