Supertron1-4B holds its own against models in the 4–8B class and surpasses Mistral 7B on all four core benchmarks despite having nearly half the parameters.
1from transformers import AutoTokenizer, AutoModelForCausalLM
2import torch
3
4model_id = "surpem/supertron1-4b"
5
6tokenizer = AutoTokenizer.from_pretrained(model_id)
7model = AutoModelForCausalLM.from_pretrained(
8 model_id,
9 torch_dtype=torch.bfloat16,
10 device_map="auto"
11)
12
13messages = [
14 {"role": "user", "content": "Explain the difference between LoRA and full fine-tuning."}
15]
16
17text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
18inputs = tokenizer(text, return_tensors="pt").to(model.device)
19
20outputs = model.generate(**inputs, max_new_tokens=512)
21print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True))
1@misc{surpem2026supertron1,
2 title={Supertron1-4B — Efficient Instruction-Tuned Language Model},
3 author={Surpem},
4 year={2026},
5 url={https://huggingface.co/surpem/supertron1-4b},
6}
