SimpleSD-4B-thinking
This model is an example of the Simple Self-Distillation (SimpleSD) method that improves code generation by fine-tuning a language model on its own sampled outputs—without rewards, verifiers, teacher models, or reinforcement learning. Please see the paper below for more information. This uses Qwen for initialization.
- Self-distillation sampling: temperature=1.1, top_p=0.95, top_k=20
- Evaluation sampling: temperature=0.7, top_p=0.95, top_k=20
paper: https://arxiv.org/abs/2604.01193
code: https://github.com/apple/ml-ssd
Notes
- These are research checkpoints for reproducibility.
- They are not optimized Qwen releases.
- They don't represent a broader open-source model strategy.
Usage
from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained("apple/SimpleSD-4B-thinking")
tokenizer = AutoTokenizer.from_pretrained("apple/SimpleSD-4B-thinking")
Method
SimpleSD samples solutions from the base model using non-unit temperature and top-k/top-p truncation, then fine-tunes on those samples via standard supervised learning. Despite its simplicity, SimpleSD yields large gains on competitive programming benchmarks, with improvements concentrating on harder problems. The mechanism traces to resolving a precision–exploration conflict: SimpleSD reshapes token distributions in a context-dependent way so that a single global decoding configuration becomes far more effective at evaluation time.
Results
LiveCodeBench (%)
| Model | LCBv6 pass@1 | LCBv6 pass@5 | LCBv5 pass@1 | LCBv5 pass@5 |
|---|---|---|---|---|
| Qwen3-4B-Thinking-2507 (base) | 54.5 | 67.5 | 59.6 | 70.3 |
| + SimpleSD (this model) | 57.8 (+3.3) | 71.4 (+3.9) | 63.1 (+3.5) | 74.7 (+4.4) |
Paper
Embarrassingly Simple Self-Distillation Improves Code Generation
@misc{zhang2026embarrassinglysimpleselfdistillationimproves,
title={Embarrassingly Simple Self-Distillation Improves Code Generation},
author={Ruixiang Zhang and Richard He Bai and Huangjie Zheng and Navdeep Jaitly and Ronan Collobert and Yizhe Zhang},
year={2026},
eprint={2604.01193},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2604.01193},
}
License
This model is released under the Apple Machine Learning Research Model License.