Views
No views yet
| Model | AIME25 | LiveCodeBench v6 | GPQA-D | Average |
|---|---|---|---|---|
| gpt-oss-20b | 91.7 | 61.0 | 71.5 | 74.7 |
| Qwen3-30B-A3B-Thinking-2507 | 85.0 | 66.0 | 73.4 | 74.8 |
| NVIDIA-Nemotron-3-Nano-30B-A3B | 89.1 | 68.3 | 73.0 | 76.8 |
| DASD-30B-A3B-Thinking-Preview (Ours) | 86.7 | 72.8 | 72.3 | 77.3 |
Note1: To demonstrate the scalability and efficiency of our data recipe, this preview model was trained only on the first-stage (Low-Temperature) dataset (~105K samples) derived from our 4B pipeline, without any re-curation or additional RL. Even with this lightweight recipe, it achieves excellent performance among open MoE models.
Note2: This model (DASD-30B-A3B-Thinking-Preview) is a preliminary research artifact trained only on the first stage (Low-Temperature Sampling) of our pipeline to demonstrate the scalability of our data recipe. For the fully trained model and complete methodology, please refer to DASD-4B-Thinking and our Technical Report.
1from transformers import AutoModelForCausalLM, AutoTokenizer
2
3model_name = "Alibaba-Apsara/DASD-30B-A3B-Thinking-Preview"
4
5tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
6model = AutoModelForCausalLM.from_pretrained(
7 model_name,
8 torch_dtype="auto",
9 device_map="auto",
10 trust_remote_code=True,
11)
12
13prompt = "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?"
14messages = [
15 {"role": "system", "content": "You are a helpful assistant."},
16 {"role": "user", "content": prompt}
17]
18
19text = tokenizer.apply_chat_template(
20 messages,
21 tokenize=False,
22 add_generation_prompt=True,
23)
24
25model_inputs = tokenizer([text], return_tensors="pt").to(model.device)
26
27generated_ids = model.generate(
28 **model_inputs,
29 max_new_tokens=81920,
30)
31
32output_ids = generated_ids[0][len(model_inputs.input_ids[0]):].tolist()
33content = tokenizer.decode(output_ids, skip_special_tokens=True)
34print(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.
python -m sglang.launch_server --model-path Alibaba-Apsara/DASD-30B-A3B-Thinking-Preview --context-length 262144vllm serve Alibaba-Apsara/DASD-30B-A3B-Thinking-Preview --max-model-len 2621441@article{yan2026dasd,
2 title={Distribution-Aligned Sequence Distillation for Superior Long-CoT Reasoning},
3 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},
4 year={2026},
5 journal={arXiv preprint arXiv:2601.09088},
6 url={https://arxiv.org/abs/2601.09088}
7}
8
9@article{liu2025where,
10 title={Where Did This Sentence Come From? Tracing Provenance in LLM Reasoning Distillation},
11 author={Liu, Kaiyuan and Yan, Shaotian and Miao, Rui and Wang, Bing and Shen, Chen and Zhang, Jun and Ye, Jieping},
12 journal={arXiv preprint arXiv:2512.20908},
13 year={2025}
14}