You can deploy this model and start earning money today!

FlyCockpit/
DeepSeek-V4-Flash-0731-vision

Name: DeepSeek-V4-Flash-0731-vision
Availability: InStock

deepseek-ai/DeepSeek-V4-Flash-0731

finetune

apache-2.0

endpoints_compatible

Views

No views yet

DeepSeek-V4-Flash — Vision Encoder

Give your self-hosted DeepSeek eyes.
Official production tower + projector for DeepSeek-V4-Flash-0731.

This repo is only the vision stack (about 900 MB total). The language model itself is still the upstream 0731 FP8 checkpoint (about 167 GB).

text

image  →  DeepEncoderV2 tower  →  projector  →  splice at token 129279
                                              →  DeepSeek-V4-Flash-0731

Serve it on 2× DGX Spark:
→ FlyCockpit/DeepSeek-V4-Vision-2x-DGX-Sparks

Download

bash

hf download FlyCockpit/DeepSeek-V4-Flash-0731-vision \
  --local-dir $HOME/dsv4-vision-assets

Or pull just the two weight files:

bash

hf download FlyCockpit/DeepSeek-V4-Flash-0731-vision \
  adapter/latest.pt \
  tower/deepencoder_v2_tower.safetensors \
  --local-dir $HOME/dsv4-vision-assets

Files

path	what	size	md5
`adapter/latest.pt`	use this — final projector	40 MB	`d9b3b3bda8f790ecf7cd5a98e6fb93a5`
`adapter/merged-004800-5af0c5.pt`	same bytes (named by step)	40 MB	`d9b3b3bda8f790ecf7cd5a98e6fb93a5`
`tower/deepencoder_v2_tower.safetensors`	frozen image tower	865 MB	`2d5dba626d816cc367d28b32e744830e`
`config.json`	release metadata	tiny	—

Final training step: 4800. Layout: tiles=2 (257 / 769 / 1281 image tokens).

What it's for

Strength	Notes
UI / screenshot understanding	Primary target
Documents & on-screen text	Strong
Everyday photos	Decent
Click agents on arbitrary real UIs	Not claimed — see below

Synthetic GUI coordinate probes pass; real-page computer-use does not transfer cleanly yet. Full write-up lives in the playbook's capabilities doc.

Compatibility (important)

This encoder maps into 0731's embedding space.

base weights	works?
`deepseek-ai/DeepSeek-V4-Flash-0731`	yes
older Flash / DSpark / NVIDIA NVFP4 of the old line	no

Precision (FP8 vs a future 0731-matched quant) is fine. Different training lineage is not.

Serve

Use the open playbook (Mia-style, worker-first TP=2):

https://github.com/FlyCockpit/DeepSeek-V4-Vision-2x-DGX-Sparks

bash

git clone https://github.com/FlyCockpit/DeepSeek-V4-Vision-2x-DGX-Sparks.git
cd DeepSeek-V4-Vision-2x-DGX-Sparks
cp .env.vision.example .env.vision   # set fabric IPs
bash scripts/download-assets.sh
bash scripts/build-image.sh
bash scripts/start-vision.sh         # run on head

API: http://<head>:8899/v1 · model id: deepseek-v4-flash-0731-vision

License

Adapter packaging in this repo: Apache-2.0
DeepEncoderV2 tower & DeepSeek-V4-Flash-0731 backbone: upstream terms apply