Qwen3.6-35B-A3B hits local coding benchmark milestone

The era of "good enough" local coding models has ended—Qwen3.6 is legitimately competitive with top-tier closed models.

• –Sparse MoE architecture (35B total, 3B active) allows it to fit in 24GB VRAM while delivering performance that rivals dense 70B+ models.
• –Integration with OpenCode harness enables "agentic execution," moving beyond simple snippets to repository-level refactors and bug fixes.
• –Native 262k context window and "Thinking Preservation" features are game-changers for long-running iterative coding sessions.
• –Outperforms its predecessor (Qwen3.5) on SWE-bench and Terminal-Bench, specifically excelling in complex tool-calling scenarios.
• –Locally deployable via llama.cpp, it represents a significant shift for privacy-conscious developers who need frontier-level coding assistance.