Qwen3-Coder-Next leads SWE-rebench pass@5

The interesting part is not just that Qwen3-Coder-Next is good — it is that an instruct-style open model is now competitive in exactly the multi-step recovery-heavy workflows where coding agents usually break.

• –SWE-rebench’s January 2026 leaderboard explicitly calls Qwen3-Coder-Next the best open-source model by pass@5, with the site highlighting its strong showing despite only ~3B active parameters
• –Its 64.6% pass@5 beats every non-harness model on the board, even though Claude Code and Junie still rank higher as full agent systems rather than raw models
• –The result matters for developers running private coding workflows locally, where open weights, controllable inference, and lower operational risk can matter more than absolute frontier polish
• –The benchmark notes a practical catch: hosted providers often lack token or prefix caching support for Qwen3-Coder-Next, which can hurt real-world agent efficiency even when raw capability is excellent
• –This also strengthens the case that Qwen’s coding line is iterating unusually fast, with Qwen3-Coder-Next materially outperforming earlier Qwen coding and general-purpose variants on agentic software tasks