# DASD-4B-Thinking

[![GitHub](https://img.shields.io/badge/GitHub-DASD--Thinking-181717?logo=github&logoColor=white)](https://github.com/D2I-ai/dasd-thinking)

[![Hugging Face](https://img.shields.io/badge/%F0%9F%A4%97%20Checkpoint-DASD--4B--Thinking-yellow)](https://huggingface.co/Alibaba-Apsara/DASD-4B-Thinking) [![Hugging Face](https://img.shields.io/badge/%F0%9F%A4%97%20Checkpoint-DASD--30B--A3B--Thinking--Preview-yellow)](https://huggingface.co/Alibaba-Apsara/DASD-30B-A3B-Thinking-Preview) [![Hugging Face](https://img.shields.io/badge/%F0%9F%A4%97%20Dataset-Superior--Reasoning--SFT--gpt--oss--120b-red)](https://huggingface.co/datasets/Alibaba-Apsara/Superior-Reasoning-SFT-gpt-oss-120b) [![Hugging Face](https://img.shields.io/badge/%F0%9F%A4%97%20Dataset-Superior--Reasoning--SFT--gpt--oss--120b--Logprob-red)](https://huggingface.co/datasets/Alibaba-Apsara/Superior-Reasoning-SFT-gpt-oss-120b-Logprob) ## 🚀 Introduction We release **DASD-4B-Thinking**, a compact yet capable 4B dense language model specialized in **long chain-of-thought (Long-CoT) reasoning** across mathematics, code generation, and scientific reasoning. DASD-4B-Thinking is post-trained from **Qwen3-4B-Instruct-2507** (non-thinking student) and distilled from **gpt-oss-120b** (teacher) via a **distribution-aligned sequence distillation** pipeline, achieving strong long-cot reasoning performance with substantially fewer training samples (**448K**) than many existing larger models.

## 📊 Performance | Model | Data | AIME24 | AIME25 | LiveCodeBench v5 | LiveCodeBench v6 | GPQA-D | |---------------------------------|---------|--------|--------|--------|--------|--------| | Qwen3-4B-Thinking-2507 |❌ | - | 81.3 | - | 55.2 | 65.8 | | Qwen3-14B |❌ | 79.3 | 70.4 | 63.5 | - | 64.0 | | Qwen3-32B |❌ | 81.4 | 72.9 | 65.7 | - | 68.4 | | DeepSeek-R1-0528-Qwen3-8B |❌ | 86.0 | 76.3 | 60.5 | - | 61.1 | | GLM-Z1-32B-0414 |❌ | 80.8 | 63.6 | 59.1 | - | 66.1 | | GLM-Z1-9B-0414 |❌ | 76.4 | 56.6 | 51.8 | - | 58.5 | | Mistral3-3B |❌ | - | 72.1 | 54.8 | - | 53.4 | | Mistral3-8B |❌ | - | 78.7 | 61.6 | - | 66.8 | | AM-thinking-v1 |✅ | 85.3 | 74.4 | 70.3 | - | - | | POLARIS-4B-Preview |✅ | 81.2 | 79.4 | - | - | - | | OpenThoughts3-7B |✅ | 69.0 | 53.3 | 51.7 | - | 53.7 | | Pai-DistillQwen-ThoughtY-4B |✅ | 76.7 | - | - | - | 56.1 | | Pai-DistillQwen-ThoughtY-8B |✅ | 76.7 | - | - | - | 62.1 | | NVIDIA-OpenReasoning-Nemotron-7B |✅ | 84.7 | 78.2 | 63.9 | - | 61.4 | | NVIDIA-Nemotron-Ultra-253B |✅ | 80.8 | 72.5 | 68.1 | - | 76.0 | | **DASD-4B-Thinking (Ours)** |✅ | **88.5** | **83.3** | **69.3** | **67.5** | **68.4** | --- ## 💡 Why DASD-4B-Thinking Matters While the community rushes to build distilled reasoning model using massive datasets (often millions of samples), DASD-4B-Thinking proves that *distribution alignment matters more than data quantity*. It establishes a new baseline for **data-efficient distillation**, delivering flagship-level reasoning in a 4B model that can run on consumer hardware. DASD-4B-Thinking democratizes the training recipe: * **Open-Source Model**: It achieves State-of-the-Art performance among open-source models of comparable scale and outperforms significantly larger models. * **Extreme Data Efficiency**: Achieves these results using only **448K training samples**, an order of magnitude fewer than comparable efforts. * **Novel pipeline**: It presents a systematic reexamining of sequence-level distillation and introduces **a novel distribution-aligned sequence distillation pipeline**. * **Open-Source Data**: We release the [Alibaba-Apsara/Superior-Reasoning-SFT-gpt-oss-120b](https://huggingface.co/datasets/Alibaba-Apsara/Superior-Reasoning-SFT-gpt-oss-120b), allowing the community to reproduce our off-policy **temperature-scheduled** pipeline: * **105K** Low-Temperature responses for stability (Stage 1). * **330K** High-Temperature responses for diversity (Stage 2). * **Proven Scalability**: The exact same data recipe generalizes effectively to larger architectures, as demonstrated by our **[DASD-30B-A3B-Thinking-Preview](https://huggingface.co/Alibaba-Apsara/DASD-30B-A3B-Thinking-Preview)** (MoE), which achieves competitive performance without extra RL. ## ⚙️ Post-Training Pipeline DASD-Thinking introduces a new paradigm of **Distribution-Aligned Sequence Distillation**. This represents an enhanced sequence-level distillation pipeline that incorporates **Temperature-scheduled Learning**, **Divergence-aware Sampling**, and **Mixed-policy Distillation** , achieving efficient capability transfer with a minimal amount of data (**448K**). Please refer to our [report](https://arxiv.org/abs/2601.09088) for more details.

## ⚡ Quick Start ```python from transformers import AutoModelForCausalLM, AutoTokenizer model_name = "Alibaba-Apsara/DASD-4B-Thinking" tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True) model = AutoModelForCausalLM.from_pretrained( model_name, torch_dtype="auto", device_map="auto", trust_remote_code=True, ) prompt = "Natalia sold clips to 48 of her friends in April, and then she sold half as many clips in May. How many clips did Natalia sell altogether in April and May?" messages = [ {"role": "system", "content": "You are a helpful assistant."}, {"role": "user", "content": prompt} ] text = tokenizer.apply_chat_template( messages, tokenize=False, add_generation_prompt=True, ) model_inputs = tokenizer([text], return_tensors="pt").to(model.device) generated_ids = model.generate( **model_inputs, max_new_tokens=81920, ) output_ids = generated_ids[0][len(model_inputs.input_ids[0]):].tolist() content = tokenizer.decode(output_ids, skip_special_tokens=True) print(content) ``` > Note: We include the system prompt, as it was used during all training stages. To ensure consistent output quality, we recommend including the same system prompt during actual usage; otherwise, the model's responses may be affected. For deployment, you can use sglang>=0.4.6.post1 or vllm>=0.8.5 or to create an OpenAI-compatible API endpoint: - SGLang: ``` python -m sglang.launch_server --model-path Alibaba-Apsara/DASD-4B-Thinking --context-length 262144 ``` - vLLM: ``` vllm serve Alibaba-Apsara/DASD-4B-Thinking --max-model-len 262144 ``` ## 💡Best Practices To achieve optimal performance, we suggest using **Temperature=1.0, TopP=1.0**. ## 📜 Licence The model weights are licensed under Apache 2.0 License. ## ⚠️ Limitation While DASD-4B-Thinking demonstrates remarkable performance across mathematical, scientific, and coding benchmarks, **it is currently limited by the absence of tool integration and function calling capabilities.** Operating strictly within the text space, the model cannot interact with external interfaces such as code executors or APIs, which constrains its utility in agent-based workflows; however, future iterations aim to bridge this gap by integrating capabilities like knowledge retrieval and tool invocation to support more complex, interactive reasoning tasks. ## 📚 Citation DASD-Thinking is developed by Alibaba Cloud, as part of our mission to advance open, efficient, and trustworthy reasoning systems. If you find this work useful in your research or applications, please cite our technical report. ```bibtex @article{yan2026dasd, title={Distribution-Aligned Sequence Distillation for Superior Long-CoT Reasoning}, author={Yan, Shaotian and Liu, Kaiyuan and Shen, Chen and Wang, Bing and Fan, Sinan and Zhang, Jun and Wu, Yue and Wang, Zheng and Ye, Jieping}, year={2026}, journal={arXiv preprint arXiv:2601.09088}, url={https://arxiv.org/abs/2601.09088} } @article{liu2025where, title={Where Did This Sentence Come From? Tracing Provenance in LLM Reasoning Distillation}, author={Liu, Kaiyuan and Yan, Shaotian and Miao, Rui and Wang, Bing and Shen, Chen and Zhang, Jun and Ye, Jieping}, journal={arXiv preprint arXiv:2512.20908}, year={2025} } ``` We welcome collaboration, feedback, and community contributions to push the boundaries of what small models can reason about—transparently and responsibly.