(논문 요약) Qwen3-Coder: Agentic Coding in the World (blog)

핵심 내용

• Architecture: MoE
  • Total Params: 480B
  • Active Params: 35B
• Contex length: 256K

학습

• 7.5T tokens (70% code ratio)
• Qwen2.5-Coder 로 noisy data 를 clean 및 rewrite 하여 data quality 높임
• Post training: execution-driven large-scale reinforcement learning
  • real world coding tasks 의 test cases 를 점진적으로 복잡하게 발전시켜나감
  • “Hard to Solve, Easy to Verify”

성능