Qwen3-1.7B tops sub-2B agent race

The real bottleneck under 2B isn't raw intelligence, it's orchestration: tiny models can call tools, but keeping them stable across a loop is the hard part. Qwen3 is one of the few open-weight families explicitly built around agent capabilities and external-tool integration. The 1.7B checkpoint keeps a 32K context window and a thinking/non-thinking switch, which matters in multi-turn workflows. The Reddit skepticism matches the broader small-model picture: below the 1-3B band, brittle tool calls and state drift become the main failure mode. If you only need narrow function calling, a task-specific 1.5B fine-tune may beat the base model, but Qwen3-1.7B is still the broadest default.