Qwen3-Coder-Next Technical Report

Описание: Qwen3-Coder-Next is an open-weight language model introduced as a specialized solution for coding agents. With a total of 80 billion parameters, it uniquely activates only 3 billion during inference, achieving strong coding capabilities with remarkable efficiency. The core innovation lies in its scalable agentic training approach, which involves large-scale synthesis of verifiable coding tasks paired with executable environments. This method allows the model to learn directly from environment feedback through mid-training adaptation and reinforcement learning. Qwen3-Coder-Next demonstrates competitive performance on agent-centric benchmarks like SWE-Bench Pro and Terminal-Bench, often outperforming models with significantly larger active parameter counts. Its staged training pipeline includes continued pretraining, supervised fine-tuning, and expert specialization. This efficiency makes it ideal for production coding agents where latency and cost are critical constraints, suggesting that advanced agentic training is a key driver for real-world coding agent development.
#Qwen3CoderNext #LLM #CodingAgent #AI #OpenWeightModel #AgenticTraining #MachineLearning #SoftwareEngineering #CodeGeneration #EfficientAI

paper - https://github.com/QwenLM/Qwen3-Coder...
subscribe - https://t.me/arxivpaper
donations:
USDT: 0xAA7B976c6A9A7ccC97A3B55B7fb353b6Cc8D1ef7
BTC: bc1q8972egrt38f5ye5klv3yye0996k2jjsz2zthpr
ETH: 0xAA7B976c6A9A7ccC97A3B55B7fb353b6Cc8D1ef7
SOL: DXnz1nd6oVm7evDJk25Z2wFSstEH8mcA1dzWDCVjUj9e
created with NotebookLM