Skip to main content

DeepSeek-VL2 (Open-Weight)

DeepSeek-VL2 is DeepSeek's open-weight multimodal model family on Hugging Face. No first-party hosted API exists for it -- DeepSeek's own Platform API (deepseek-flash) only serves deepseek-flash / deepseek-v4-pro, so this local adapter is the only way to use DeepSeek-VL2 today.

Unlike this repo's other local VLMs (Kimi-VL, Qwen3.8), DeepSeek-VL2 does not load via plain trust_remote_code=True + AutoProcessor -- it needs DeepSeek's own deepseek_vl2 Python package (not on PyPI) for the processor and <|User|>/<|Assistant|> conversation format.

Overview​

FeatureValue
Default modeldeepseek-ai/deepseek-vl2-tiny
ArchitectureMoE VLM, 27B total base, per-variant activated params below
LicenseCode: MIT. Weights: DeepSeek Model License (commercial use permitted)
CapabilitiesImage, multi-image, and (via frame sampling) video understanding
Min VRAMReported figures vary a lot by source -- single-GPU friendly for tiny; budget 40GB+ for small, more for the full model. Check the model card before picking a variant.
ModelTotal / ActivatedUse Case
deepseek-ai/deepseek-vl2-tiny (default)3.37B / 1.0B activePractical single-GPU default
deepseek-ai/deepseek-vl2-small16.1B / 2.8B activeBetter quality, more VRAM
deepseek-ai/deepseek-vl2 ("base")27B / 4.5B activeHeaviest, best quality

Installation​

pip install opentryon[local]   # torch, transformers, etc.
pip install "git+https://github.com/deepseek-ai/DeepSeek-VL2.git"   # deepseek_vl2 (not on PyPI)
pip install decord             # only needed for understand_video()
note

DeepSeek-VL2's model card documents transformers==4.38.2. The local extra pins a different version shared across all local models. If you hit loading errors, install the exact version DeepSeek recommends in a dedicated environment.

Quick Start​

Image Understanding​

from tryon.models.deepseek_vl2 import DeepSeekVL2Adapter

adapter = DeepSeekVL2Adapter()  # downloads deepseek-vl2-tiny on first use

result = adapter.understand_image(
    "garment.jpg",
    prompt="Describe this outfit: color, pattern, style, fit, and material.",
)
print(result["text"])

Video Understanding​

DeepSeek-VL2 has no official video recipe -- video is handled the same way as Kimi-VL/Qwen3.8 local: uniformly sample frames and pass them as a multi-image prompt.

result = adapter.understand_video(
    "runway_clip.mp4",
    prompt="Summarize the styling shown in this video.",
    num_frames=8,
)
print(result["text"])

API Reference​

DeepSeekVL2Adapter​

class DeepSeekVL2Adapter:
    def __init__(
        self,
        model_id: Optional[str] = None,   # DEEPSEEK_VL2_MODEL_ID or deepseek-ai/deepseek-vl2-tiny
        device: Optional[str] = None,     # defaults to CUDA if available, else CPU
        torch_dtype: str = "auto",        # bf16 on CUDA
    )

understand_image / understand_video / understand​

def understand_image(
    self, image, prompt="Describe the content of the image in detail.",
    max_new_tokens=512, do_sample=False, temperature=0.8,
) -> Dict[str, Any]  # {"text": ..., "model": ...}

def understand_video(
    self, video, prompt="Describe what happens in this video.",
    num_frames=8, max_new_tokens=512, do_sample=False, temperature=0.8,
) -> Dict[str, Any]

def understand(
    self, image=None, video=None, prompt="Describe the content in detail.",
    num_frames=8, max_new_tokens=512, do_sample=False, temperature=0.8,
) -> Dict[str, Any]

understand is the single entry point that accepts image and/or video (at least one required) -- this is what opentryon understand --model deepseek-vl2 calls.

Using the opentryon CLI​

opentryon understand --model deepseek-vl2 --image garment.jpg \
  --prompt "Describe this outfit."

opentryon understand --model deepseek-vl2 --video runway_clip.mp4 --num-frames 12

The CLI checks that torch is installed before attempting to load the model (pip install opentryon[local] hint if missing) -- it does not check for the separate deepseek_vl2 package, which raises its own ImportError with the git install command if missing.

Choosing a Model Variant​

adapter = DeepSeekVL2Adapter(model_id="deepseek-ai/deepseek-vl2-small")
# or: export DEEPSEEK_VL2_MODEL_ID=deepseek-ai/deepseek-vl2-small

Troubleshooting​

ImportError: DeepSeek-VL2 needs DeepSeek's own 'deepseek_vl2' package​

  • Install it from GitHub: pip install "git+https://github.com/deepseek-ai/DeepSeek-VL2.git". It is intentionally not part of opentryon[local] (not on PyPI, maintained by DeepSeek).

ImportError for decord​

  • Video understanding needs frame sampling: pip install decord.

trust_remote_code / loading errors​

  • Try DeepSeek's exact pinned transformers==4.38.2 in a dedicated environment (see note above).

References​