š Introduction
We release AM-Thinkingāv1, a 32B dense language model focused on enhancing reasoning capabilities.
Built on QwenāÆ2.5ā32BāBase, AM-Thinkingāv1 shows strong performance on reasoning benchmarks, comparable to much larger MoE models like DeepSeekāR1, Qwen3ā235BāA22B, Seed1.5-Thinking, and larger dense model like Nemotron-Ultra-253B-v1.
š§© Why Another 32B Reasoning Model Matters?
Large MixtureāofāExperts (MoE) models such as DeepSeekāR1 or Qwen3ā235BāA22B dominate leaderboardsābut they also demand clusters of highāend GPUs. Many teams just need the best dense model that fits on a single card.
AMāThinkingāv1 fills that gap while remaining fully based on open-source components:
- Outperforms DeepSeekāR1 on AIMEā24/ā25 & LiveCodeBench and approaches Qwen3ā235BāA22B despite being 1/7āth the parameter count.
- Built on the publicly availableāÆQwenāÆ2.5ā32BāBase, as well as the RL training queries.
- Shows that with a wellādesigned postātraining pipeline ( SFT + dualāstage RL ) you can squeeze flagshipālevel reasoning out of a 32āÆB dense model.
- Deploys on one A100ā80āÆGB with deterministic latencyāno MoE routing overhead.
š ļø Use Cases
1) Code Generation
2) Logic
3) Writing
ā”Ā Quick start
from transformers import AutoModelForCausalLM, AutoTokenizerĀ model_name = "a-m-team/AM-Thinking-v1"Ā tokenizer = AutoTokenizer.from_pretrained(model_name)model = AutoModelForCausalLM.from_pretrained( model_name, torch_dtype="auto", device_map="auto")Ā prompt = "How can I find inner peace?"messages = [ {"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)Ā generated_ids = model.generate( **model_inputs, max_new_tokens=49152)output_ids = generated_ids[0][len(model_inputs.input_ids[0]):].tolist() Ā response = tokenizer.decode(output_ids, skip_special_tokens=True)think_content = response.split("<think>")[1].split("</think>")[0]answer_content = response.split("<answer>")[1].split("</answer>")[0]Ā print (f"user prompt: {prompt}")print (f"model thinking: {think_content}")print (f"model answer: {answer_content}")
Note: We have included the system prompt in the tokenizer configuration, as it was used during both the SFT and RL stages. To ensure consistent output quality, we recommend including the same system prompt during actual usage; otherwise, the model's responses may be significantly affected.
Quantized versions for compact devices
A series of quantized versions for AM-Thinking-v1 model.
For use with llama.cpp and Ollama
is available at AM-Thinking-v1-gguf.
š§ Post-training pipeline
To achieve its strong reasoning ability, AMāThinkingāv1 goes through a carefully designed post-training pipeline.
Below we describe the key stages involved in turning a base model into a high-performing reasoner:
StepāÆ1Ā ā Coldāstart SFT.
We begin with the open-sourced QwenāÆ2.5ā32BāBase and run a broad supervised fineātune on a blended training dataset of math, code and openādomain chat. This endows the model with a "thinkāthenāanswer" behavioural pattern and equips it with an initial capacity for reasoning.
StepāÆ2Ā ā Passārateāaware data curation.
Before any RL, the SFT model is evaluated on every mathā and codeāoriented training query. For each item we log a pass rate; only those with 0āÆ<āÆpassārateāÆ<āÆ1 are kept. In effect we discard problems the model already masters and those it utterly fails, concentrating learning on genuinely informative cases.
StepāÆ3Ā ā Reinforcement learningĀ .
We adopt a twoāstage GRPO scheme: StageāÆ1 trains only on math and code queries. Once it converges, stage 2 starts by removing every query the model answered 100% correctly in StageāÆ1 and adjusting key hyperāparameters such as maximum generation length and learning rate.
ā ļø Limitations
While AMāThinkingāv1 excels at pure language reasoning and openādomain chat, it has not yet been trained for structured functionācalling or toolāuse workflows, which restricts its usefulness in agentāstyle applications that must act on external systems.
Improving the model's ability to follow complex instructions is also an important direction for our future work.
In addition, our safety alignment is still at an early stage, so more rigorous redāteaming are required to reduce potential harms.
š Citation
The a-m-team is an internal team at Beike (Ke.com), dedicated to exploring AGI technology.
If you find our work helpful, feel free to give us a cite.
@misc{ji2025amthinkingv1advancingfrontierreasoning, title={AM-Thinking-v1: Advancing the Frontier of Reasoning at 32B Scale}, author={Yunjie Ji and Xiaoyu Tian and Sitong Zhao and Haotian Wang and Shuaiting Chen and Yiping Peng and Han Zhao and Xiangang Li}, year={2025}, eprint={2505.08311}, archivePrefix={arXiv}, primaryClass={cs.CL}, url={https://arxiv.org/abs/2505.08311}, }