Supertron3-0.8B: Edge Foundation Model for Tool Calling and Computer Use Agents
Model Description
Supertron3-0.8B is a compact Vision-Language Model (VLM) purpose-built for GUI Agents and agentic tool calling at the edge. It operates across diverse digital environments — web, desktop, and CLI — by interpreting visual interfaces, reasoning over complex content, and emitting precise actions (pyautogui-style computer use) or valid JSON function calls.
At a 1.7GB footprint, Supertron3-0.8B delivers competitive agentic performance for low-latency, on-device deployment.
- Developed by: Suprem Org
- Model type: Vision-Language Model for Navigation, Computer Use, and Tool Calling Agents
- Architecture: Hybrid Gated DeltaNet + Attention (24 layers, 1024 hidden), 0.8B params, Vision Encoder, 262K native context
- Fine-tuned from model:
Qwen/Qwen3.5-0.8B - License: Apache 2.0
Get Started
Transformers
from transformers import AutoProcessor, AutoModelForImageTextToText
import torch
model = AutoModelForImageTextToText.from_pretrained(
"Surpem/Supertron3-0.8B", trust_remote_code=True,
torch_dtype=torch.bfloat16, device_map="auto"
)
processor = AutoProcessor.from_pretrained("Surpem/Supertron3-0.8B", trust_remote_code=True)
messages = [
{"role":"system","content": "You are Supertron3, precise tool caller. Output ONLY JSON array of tool calls.\nAvailable tools:\n[{\"name\":\"get_weather\",\"description\":\"get weather\",\"parameters\":{\"properties\":{\"city\":{\"type\":\"string\"}}}}]"},
{"role":"user","content": "What's weather in Paris on 2026-09-15?"}
]
text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = processor(text=[text], return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=256, do_sample=False)
print(processor.decode(out[0][inputs.input_ids.shape[1]:], skip_special_tokens=True))
vLLM / SGLang
vllm serve Surpem/Supertron3-0.8B --dtype bfloat16 --max-model-len 8192
Results
Tool Calling & Agentic Benchmarks
Supertron3-0.8B was evaluated on BFCL-style function calling (single-call, multi-tool, nested arguments) alongside real-world web-agent and computer-use benchmarks (Mind2Web, OmniAct). Despite being the smallest model in the comparison, Supertron3 ranks first on BFCL while being the only model that can reliably act on a desktop — the base models score higher on generic tool priors but fail completely at computer use.
Table 1: Evaluation results on tool calling, web navigation, and computer-use benchmarks.
| Model | Params | BFCL ↑ | Mind2Web (step acc) ↑ | Computer Use ↑ |
|---|---|---|---|---|
| Supertron3-0.8B (ours) | 0.8B | 82% | 77% | 100% |
| Qwen3.5-0.8B (base) | 0.8B | 56% | 80% | 0% |
| North Micro Vision Instruct | ~2B | 61% | — | — |
| Qwen3.5-4B | 4B | 69% | — | — |
Computer Use & Grounding
Supertron3 excels at localizing UI elements and emitting executable actions — a capability entirely absent in the base model. The finetune taught the base to act, not just chat.
Limitations
- 0.8B capacity: strong single-turn tool routing and short-horizon computer use; long-horizon workflows remain open.
- Grounding ceiling: ScreenSpot-Pro-class precision requires larger vision encoders.
Citation
@misc{suprem2026supertron3,
title={Supertron3-0.8B: Edge Foundation Model for Tool Calling and Computer Use Agents},
author={Suprem},
year={2026},
url={https://huggingface.co/Surpem/Supertron3-0.8B},
}
@article{qwen35,
title={Qwen3.5 Technical Report},
year={2026}
}