Here we provide different quantized versions for the AM-Thinking-v1 model.
For use with llama.cpp and Ollama,
with smaller memory consumption.
Note w.r.t. system prompt
This model is trained with explicit system prompt:
You are a helpful assistant. To answer the user's question, you first think about the reasoning process and then provide the user with the answer. The reasoning process and answer are enclosed within <think> </think> and <answer> </answer> tags, respectively, i.e., <think> reasoning process here </think> <answer> answer here </answer>.
When using the model, please use this system prompt, otherwise the model could have undefined behavior.
In the Ollama version the system prompt and chat template (we use vanilla Qwen2.5 chat template) is already integrated into the Modelfile so just use as is.
For the llama.cpp version an explicit -sys shall be used to add system prompt.
Use with llama.cpp
Download one of the model files, e.g. AM-Thinking-v1.Q4_K_M.gguf, then use llama.cpp to chat or serve.
Make sure to use with explicit system prompt, for example:
llama-cli -m AM-Thinking-v1.Q4_K_M.gguf -cnv -sys "You are a helpful assistant. To answer the user's question, you first think about the reasoning process and then provide the user with the answer. The reasoning process and answer are enclosed within <think> </think> and <answer> </answer> tags, respectively, i.e., <think> reasoning process here </think> <answer> answer here </answer>."
Use with Ollama
The Q4_K_M version has already been uploaded to Ollama.
On a Apple Macbook Pro with M3-Max, using the AM-Thinking-v1.Q4_K_M.gguf version,
we get a roughly 16 tokens/s output throughput for short prompts.
🚀 Introduction
We release AM-Thinking‑v1, a 32B dense language model focused on enhancing reasoning capabilities.
Built on Qwen 2.5‑32B‑Base, AM-Thinking‑v1 shows strong performance on reasoning benchmarks, comparable to much larger MoE models like DeepSeek‑R1, Qwen3‑235B‑A22B, Seed1.5-Thinking, and larger dense model like Nemotron-Ultra-253B-v1.
benchmark
🧩 Why Another 32B Reasoning Model Matters?
Large Mixture‑of‑Experts (MoE) models such as DeepSeek‑R1 or Qwen3‑235B‑A22B dominate leaderboards—but they also demand clusters of high‑end GPUs. Many teams just need the best dense model that fits on a single card.
AM‑Thinking‑v1 fills that gap while remaining fully based on open-source components:
Outperforms DeepSeek‑R1 on AIME’24/’25 & LiveCodeBench and approaches Qwen3‑235B‑A22B despite being 1/7‑th the parameter count.
Built on the publicly available Qwen 2.5‑32B‑Base, as well as the RL training queries.
Shows that with a well‑designed post‑training pipeline ( SFT + dual‑stage RL ) you can squeeze flagship‑level reasoning out of a 32 B dense model.
Deploys on one A100‑80 GB with deterministic latency—no MoE routing overhead.
AIME 2024
LiveCodeBench
AM-Thinking-v1 achieves strong reasoning performance with significantly fewer parameters.
🛠️ Use Cases
1) Code Generation
PROMPT :
write a python script for a bouncing red ball within a triangle, make sure to handle collision detection properly. make the triangle slowly rotate. implement it in python. make sure ball stays within the triangle
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.
🔧 Post-training pipeline
To achieve its strong reasoning ability, AM‑Thinking‑v1 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:
Step 1 – Cold‑start SFT.
We begin with the open-sourced Qwen 2.5‑32B‑Base and run a broad supervised fine‑tune on a blended training dataset of math, code and open‑domain chat. This endows the model with a "think‑then‑answer" behavioural pattern and equips it with an initial capacity for reasoning.
Step 2 – Pass‑rate‑aware data curation.
Before any RL, the SFT model is evaluated on every math‑ and code‑oriented training query. For each item we log a pass rate; only those with 0 < pass‑rate < 1 are kept. In effect we discard problems the model already masters and those it utterly fails, concentrating learning on genuinely informative cases.
Step 3 – Reinforcement learning .
We adopt a two‑stage GRPO scheme: Stage 1 trains only on math and code queries. Once it converges, stage 2 starts by removing every query the model answered 100% correctly in Stage 1 and adjusting key hyper‑parameters such as maximum generation length and learning rate.
⚠️ Limitations
While AM‑Thinking‑v1 excels at pure language reasoning and open‑domain chat, it has not yet been trained for structured function‑calling or tool‑use workflows, which restricts its usefulness in agent‑style 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 red‑teaming are required to reduce potential harms.
📚 Citation
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},
}
