Qwen-Code v0.13.0 Preview with Agent Arena

🔑 Enhanced Key Takeaways

• •Qwen3-Coder-Next, powering Qwen-Code, features an 80B total parameter MoE architecture activating only 3B parameters per inference for efficient local coding agent deployment.[1][3]
• •The model achieves 44.3% on SWE-Bench Pro, rivaling models 10-20x larger, and supports 256K native context length.[1][3]
• •Qwen-Code roadmap outlines prior releases like v0.7.0 with experimental LSP support and Anthropic provider integration, building toward v0.13.0's agent arena.[4]

🛠️ Technical Deep Dive

• •Architecture: Hybrid Gated DeltaNet (linear attention for long-range dependencies) + MoE (512 experts, 10 activated per token) + Gated Attention for reasoning; 1 shared expert always active.[1][3]
• •Training: Large-scale executable task synthesis combined with reinforcement learning (RL).[1][3]
• •Model type: Causal language model with open weights license, optimized for coding agents in local environments.[1][3]

🔮 Future ImplicationsAI analysis grounded in cited sources

⏳ Timeline