Alibaba's Qwen3-Coder-Next: 80B-Param MoE Coding Model With Only 3B Active — Scores 70.6% on SWE-Bench

Summary: Alibaba released Qwen3-Coder-Next, an open-source coding model that routes each query through just 3B of its 80B parameters, achieving frontier-class coding performance at a fraction of the compute cost.

Key Facts

• Architecture: 80B total / 3B active parameters (ultra-sparse MoE); delivers up to 10× higher throughput than dense models of comparable quality
• Benchmarks: 70.6% on SWE-Bench Verified; outperforms Claude Opus 4.5 on secure code generation
• Context & Language Coverage: Up to 1M token context window; 370+ programming languages supported
• License: Apache 2.0 — free for commercial use and fine-tuning

Why It Matters

An open-source model that matches or beats closed, proprietary coding models at a fraction of the inference cost significantly lowers the barrier for teams that can't afford frontier API pricing. The ultra-sparse MoE approach — proven viable at this scale — may become the dominant architecture for specialized open-source models in 2026, directly pressuring vendors like GitHub Copilot and Cursor.

Read More

• Qwen3-Coder-Next: ultra-sparse open source model — VentureBeat
• Qwen3-Coder-Next official blog — Qwen AI