Step3 VL 10B
Step3 VL 10B is a lightweight open-source foundation model for image-text-to-text tasks. According to its official model card, it has a 10B parameter footprint and uses a unified pre-training approach on a 1.2T token multimodal corpus, followed by SFT and RL. It is released under the Apache-2.0 license and provides safetensors weights. The card states FP8 quantized weights are available and vLLM integration is officially supported.
At a glance
- Parameters
- 10.17B
- Architecture
- step_robotics
- License
- apache-2.0
- Software
- vLLM, Transformers
Will it run on your machine?
Save your machine to see a personalized rating and its reasoning.
Add my hardwareBest for
- Visual perception tasks such as grounding, counting, OCR, and GUI interactions.
- Complex reasoning involving general knowledge and education-centric tasks.
- Multimodal image-text-to-text generation.
- Research on parallel coordinated reasoning (PaCoRe) for test-time compute allocation.
- Interactive visual question answering or exploration based on the model's perception and reasoning capabilities.
Tradeoffs
- The model card mentions corrected benchmark results are pending for some evaluations, so some reported numbers may be inaccurate.
- Specific hardware requirements are not provided in the supplied data.
- Performance comparisons are based on claims in the model card and have not been independently verified.
- The model is intended for research and development; no specific real-world deployment guarantees are stated.
- Licensing details beyond Apache-2.0 (e.g., commercial use terms) are not fully described in the supplied data.
Ways to run it
vLLM · Linux
Repository-specific command in publisher documentation; confirm dependencies and hardware in the source.
vllm serve --model stepfun-ai/Step3-VL-10B -tp 1 --reasoning-parser deepseek_r1 --enable-auto-tool-choice --tool-call-parser hermes --trust-remote-codeTransformers · See official guide
Repository-specific command in publisher documentation; confirm dependencies and hardware in the source.
from transformers import AutoProcessor, AutoModelForCausalLM
key_mapping = {
"^vision_model": "model.vision_model",
r"^model(?!\.(language_model|vision_model))": "model.language_model",
"vit_large_projector": "model.vit_large_projector",
}
model_path = "stepfun-ai/Step3-VL-10B"
processor = AutoProcessor.from_pretrained(model_path, trust_remote_code=True)
messages = [
{
"role": "user",
"content": [
{"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/bee.jpg"},
{"type": "text", "text": "What's in this picture?"}
]
},
]
model = AutoModelForCausalLM.from_pretrained(
model_path,
trust_remote_code=True,
device_map="auto",
torch_dtype="auto",
key_mapping=key_mapping).eval()
inputs = processor.apply_chat_template(
messages, add_generation_prompt=True, tokenize=True,
return_dict=True, return_tensors="pt"
).to(model.device)
generate_ids = model.generate(**inputs, max_new_tokens=1024, do_sample=False)
decoded = processor.decode(generate_ids[0, inputs["input_ids"].shape[-1] :], skip_special_tokens=True)
print(decoded)Explore its uses
Keep exploring
Bonsai 2 · 27B
A compressed 27B-class reasoning model with publisher-provided GGUF packs and a dedicated llama.cpp fork for CUDA, Metal, and CPU.
DeepSeek V4.1 Flash
An open-weight model for reasoning, coding, agents, and image understanding. DeepSeek V4.1 Flash combines a 552B backbone with sparse activation and up to one million tokens of context. Choose the original checkpoint or a community uncensored variant below.
Qwen3.8
Explore 4 versions of Qwen3.8. Compare model sizes, packages, hardware requirements, and available setups.