Qwen3.5-35B-A3B wows local coders

This reads like an early sign that Qwen found the right balance for agentic coding: deep enough to investigate problems, fast enough to stay usable, and efficient enough to run locally.

• –The praise is specific, not generic hype: users call out debugging depth, better summaries, and fewer pointless thought loops.
• –Qwen’s official model card positions Qwen3.5 as a long-context, native multimodal family, which matches the kind of repo-level work people want from coding models.
• –The 35B total / 3B active MoE setup is the real story here: it makes strong output more accessible without demanding datacenter-scale inference.
• –If the community reaction holds up across harder evals, the smaller variants could become the more interesting release for everyday local use.