open-loopify-Qwen3-4B

Qwen3-4B-Base with a retrofitted loop — layers 13–21 run three times, 54 blocks of compute per token instead of 36, no new parameters — mid-trained on 1.57B tokens of open reasoning traces.

It ships as an ordinary Qwen3 checkpoint: the loop is unrolled into 54 layers, so transformers and vLLM load it with no custom code. Trained with open-loopify.

Results

Zero-shot, chat template, temperature 0.6, top-p 0.95, up to 16k generated tokens; AIME, HMMT and AMC averaged over 8 samples per problem. All models scored the same way, with open-loopify's evaluation script.

model	AIME 2024	AIME 2025	HMMT Feb 2025	AMC 2023	MATH-500	GPQA-Diamond
Qwen3-4B-Base	9.6	5.0	0.4	42.5	68.6	33.8
open-loopify-Qwen3-4B	42.5	36.7	22.5	80.3	90.4	37.9
Qwen3-4B (official, with large-scale RL)	62.1	47.1	32.9	88.4	94.0	53.5

The model reasons at length; with a 28k-token budget it reaches 50.0 on AIME 2024, 41.2 on AIME 2025 and 24.6 on HMMT Feb 2025.

Use

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "tyzhu/open-loopify-Qwen3-4B"
tok = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.bfloat16).cuda()

question = "What is the sum of all positive divisors of 36?"
prompt = f"{question}\nPlease reason step by step, and put your final answer within \\boxed{{}}."
inputs = tok.apply_chat_template([{"role": "user", "content": prompt}], add_generation_prompt=True,
                                 return_dict=True, return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=16384)  # samples at temperature 0.6, top-p 0.95 by default
print(tok.decode(out[0, inputs["input_ids"].shape[1]:], skip_special_tokens=True))

Or serve it: vllm serve tyzhu/open-loopify-Qwen3-4B.

Give it room to think: answers often run 5k–15k tokens.

Notes

• Use it as shipped. The loop is baked in at three passes, the count it was trained with. Rebuilding it with fewer passes breaks the model (with one pass, AIME 2024 drops below 1%).
• Training: 3000 steps × 524,288 tokens at 16k context, AdamW, learning rate 3e-5 with cosine decay, on OpenThoughts3-1.2M (traces cut off by their generator's 16k limit removed) and OpenR1-Math-220k, 2:1 in tokens.
• Not trained with reinforcement learning; no safety tuning beyond the base model's.

License

Apache-2.0, as Qwen3-4B-Base and the training data.