AM-Thinking-v1: Advancing the Frontier of Reasoning at 32B Scale

TL;DR

AM-Thinking-v1 demonstrates that a $32B$ dense LLM, built on publicly available data and a carefully designed post-training pipeline, can achieve state-of-the-art reasoning among dense models and rival larger MoE systems on math and coding benchmarks. The approach combines extensive data curation, ground-truth verification, and a two-stage SFT+RL training regime with difficulty-aware query selection and GRPO optimization. Results on AIME2024 ($85.3$), AIME2025 ($74.4$), and LiveCodeBench ($70.3$) show strong reasoning and coding capabilities with practical deployability. While maintaining open-source accessibility, the work acknowledges gaps in tool usage, multimodal support, and safety alignment, highlighting a promising direction for mid-scale, collaborative AI development.

Abstract

We present AM-Thinking-v1, a 32B dense language model that advances the frontier of reasoning, embodying the collaborative spirit of open-source innovation. Outperforming DeepSeek-R1 and rivaling leading Mixture-of-Experts (MoE) models like Qwen3-235B-A22B and Seed1.5-Thinking, AM-Thinking-v1 achieves impressive scores of 85.3 on AIME 2024, 74.4 on AIME 2025, and 70.3 on LiveCodeBench, showcasing state-of-the-art mathematical and coding capabilities among open-source models of similar scale. Built entirely from the open-source Qwen2.5-32B base model and publicly available queries, AM-Thinking-v1 leverages a meticulously crafted post-training pipeline - combining supervised fine-tuning and reinforcement learning - to deliver exceptional reasoning capabilities. This work demonstrates that the open-source community can achieve high performance at the 32B scale, a practical sweet spot for deployment and fine-tuning. By striking a balance between top-tier performance and real-world usability, we hope AM-Thinking-v1 inspires further collaborative efforts to harness mid-scale models, pushing reasoning boundaries while keeping accessibility at the core of innovation. We have open-sourced our model on \href{https://huggingface.co/a-m-team/AM-Thinking-v1}{Hugging Face}.

Figures (8)

• Figure 1: Comparison of Model Performance on Reasoning Benchmarks
• Figure 2: Method Call And Standard Input/Output test case examples
• Figure 3: Validator Input Example
• Figure 4: Instance Level Distribution (left) and Token Level Distribution (right) during SFT. It is worth noting that the proportions are computed over responses, not queries, since a single query can correspond to multiple responses in our training set.
• Figure 5: Detached Rollout and Upgrade with Streaming Load Balancing Architecture