QwQ 32B

QwQ-32B performs step-by-step mathematical problem-solving through a two-stage reinforcement learning pipeline: Stage 1 trains on math/coding tasks using outcome-based rewards from accuracy verifiers, while Stage 2 applies a general reward model to preserve instruction-following capabilities. The reasoning process is visible in output tokens, allowing users to inspect the model's intermediate steps and logical progression before the final answer, enabling verification and debugging of mathematical derivations.

QwQ-32B generates code solutions and validates them through Stage 1 RL training using code execution servers that run generated code against test cases and provide outcome-based rewards. The model learns to produce executable code that passes validation checks, with the reasoning process visible in output tokens showing problem decomposition, implementation strategy, and test case consideration before the final code output.

QwQ-32B integrates tool-use capabilities trained through Stage 2 RL using a general reward model and rule-based verifiers for agent actions. The model learns to select appropriate tools, construct valid function calls, and adapt subsequent actions based on environmental feedback from tool execution, with the reasoning process showing tool selection rationale and adaptation strategy in output tokens.

QwQ-32B undergoes Stage 2 RL training using a general reward model to align with human preferences and instruction-following requirements, preventing performance degradation in non-reasoning tasks after math/coding optimization. The model learns to follow complex multi-step instructions, maintain context across conversations, and balance reasoning transparency with practical task completion through reward signals from preference-aligned verifiers.

QwQ-32B is deployable on a single GPU through native Hugging Face Transformers integration using `AutoModelForCausalLM` and `AutoTokenizer`, with model weights available on Hugging Face Hub and ModelScope. The deployment pattern supports local inference without cloud API dependencies, enabling private reasoning workloads and custom integration into applications through standard PyTorch model loading and generation APIs.

QwQ-32B is available through Alibaba Cloud's DashScope API, providing managed inference without local GPU requirements. The API abstracts deployment complexity and provides scalable, pay-per-use access to the model with standard REST/streaming endpoints, enabling integration into applications without infrastructure management while maintaining the same reasoning and tool-use capabilities as self-hosted deployment.

QwQ-32B is accessible through Qwen Chat, a web-based interface providing browser-based access to the model without local installation or API integration. Users interact through a conversational chat interface that displays reasoning tokens and responses, enabling exploration of the model's capabilities without technical setup while maintaining the same reasoning transparency as programmatic access.

QwQ-32B is distributed under Apache 2.0 license with full model weights publicly available on Hugging Face and ModelScope, enabling unrestricted commercial use, modification, and redistribution. The open-weight distribution allows organizations to build proprietary applications, fine-tune for specific domains, and maintain full control over model deployment without licensing restrictions or usage reporting requirements.

QwQ-32B achieves 79.5% accuracy on AIME 2024 (American Invitational Mathematics Examination) and 96.4% on MATH-500, demonstrating strong mathematical reasoning capability at 32B parameters. These benchmarks measure complex multi-step mathematical problem-solving with explicit reasoning visible in output tokens, enabling evaluation of both correctness and reasoning quality without hidden reasoning processes.

QwQ-32B is evaluated on LiveCodeBench, a competitive programming benchmark measuring code generation quality through execution against test cases. The model's performance on this benchmark reflects its ability to generate correct, executable code with reasoning visible in output tokens, enabling assessment of both code quality and problem-solving approach in algorithmic contexts.