zeromodels/deepseek_vl_7b_base

See our collection for all versions of DeepSeek-VL.

Run DeepSeek-VL with Keras 3: JAX, PyTorch, or TensorFlow

kerasformers/deepseek_vl_7b_base

Paper: DeepSeek-VL: Towards Real-World Vision-Language Understanding (arXiv:2403.05525) · HF Papers

DeepSeek-VL 7B hybrid uses a dual vision tower (SigLIP @384 + SAM-style @1024) and a 3-way aligner for finer detail and small text. Import from kerasformers.models.deepseek_vl_hybrid (not deepseek_vl).

For more details on the model, please go to the upstream model card.

Pure-Keras 3 conversion of deepseek-ai/deepseek-vl-7b-base for kerasformers. One implementation runs unmodified on TensorFlow / Torch / JAX.

This is a vision-language (base) checkpoint (DeepseekVLHybridConditionalGenerate, 7B base (SigLIP + SAM hybrid)).

✨ Quick start

import os
os.environ["KERAS_BACKEND"] = "torch"  # or "jax" / "tensorflow"

from PIL import Image
from kerasformers.models.deepseek_vl_hybrid import (
    DeepseekVLHybridConditionalGenerate,
    DeepseekVLHybridProcessor,
)

model = DeepseekVLHybridConditionalGenerate.from_weights("kerasformers/deepseek_vl_7b_base")
processor = DeepseekVLHybridProcessor.from_weights("kerasformers/deepseek_vl_7b_base")

image = Image.open("your_image.jpg")
inputs = processor(
    conversation=[
        {
            "role": "user",
            "content": [
                {"type": "image", "image": image},
                {"type": "text", "text": "Describe this image in one sentence."},
            ],
        }
    ]
)
outputs = model.generate(**inputs, max_new_tokens=64)
print(processor.decode(outputs[0]))

Load any DeepSeek-VL variant the same way with from_weights("kerasformers/<variant>"):

Variant	Hub	Notes
deepseek_vl_1.3b_base	kerasformers/deepseek_vl_1.3b_base	1.3B base
deepseek_vl_1.3b_chat	kerasformers/deepseek_vl_1.3b_chat	1.3B chat
deepseek_vl_7b_base	kerasformers/deepseek_vl_7b_base	7B hybrid base
deepseek_vl_7b_chat	kerasformers/deepseek_vl_7b_chat	7B hybrid chat

Tips

• Set KERAS_BACKEND before importing Keras / kerasformers.
• Prefer Processor.from_weights(...) so vision + tokenizer match.
• 7B hybrid is a different package than 1.3B; do not mix imports.
• See docs and Loading Weights.
• Community / upstream safetensors still work via the hf: prefix, e.g. DeepseekVLHybridConditionalGenerate.from_weights("hf:deepseek-ai/deepseek-vl-7b-base").

Special Thanks

A huge thank you to the DeepSeek-VL authors for creating and releasing these models.

License: DeepSeek (other / deepseek). See LICENSE.