euhidaman/embervlm-small-trial

EMBERVLM-SMALL: Small (~137M parameters)

EmberVLM is an efficient vision-language model optimized for edge deployment and robotic applications.

Model Details

• Model Type: Vision-Language Model (VLM)
• Size: Small (~137M parameters)
• Total Parameters: 164,203,841
• Trainable Parameters: 35,943,041
• Carbon Emissions: 0.0308 kg CO2eq

Architecture

• Vision Encoder: dinov2_small
• Language Model: SmolLM-135M (135M params)
• Training Stages: 4-stage curriculum
  1. Visual-Language Alignment
  2. Multimodal Instruction Tuning
  3. Robot Fleet Selection
  4. Chain-of-Thought Reasoning

Usage

from embervlm import EmberVLM
from transformers import AutoTokenizer
from PIL import Image

# Load model and tokenizer
model = EmberVLM.from_pretrained("embervlm-small")
tokenizer = AutoTokenizer.from_pretrained("embervlm-small")

# Prepare input
image = Image.open("robot_scene.jpg")
prompt = "<image>What is happening in this scene?"

# Generate response
outputs = model.generate(image=image, prompt=prompt, tokenizer=tokenizer)
print(outputs)

Training Configuration

• Vision Backbone: dinov2_small
• Language Backbone: smollm_135m
• Optimization: AdamW with cosine learning rate schedule
• Mixed Precision: bfloat16
• Stages Completed: 1-4 (Full curriculum)

Intended Use

• Edge deployment on resource-constrained devices
• Robotic vision-language understanding
• Real-time multimodal reasoning
• Robot fleet selection and task planning

Limitations

• Optimized for efficiency over maximum accuracy
• Best suited for edge/mobile deployment scenarios
• Training focused on robot-centric scenarios

Citation

@software{embervlm_embervlm_small,
  title = {EmberVLM-Small},
  author = {EmberVLM Team},
  year = {2026},
  url = {https://huggingface.co/embervlm-small}
}