Developer ditches Cursor for Qwen 3.6 35B on 700k-line codebase

A developer working 60 hours a week on a 500,000–700,000 line enterprise software suite has replaced Cursor's Composer 2 with Qwen 3.6 35B-A3B, citing superior performance on real-world coding tasks and a fraction of the cost. Running the model on OpenRouter at an effective rate of $0.08 per million tokens after prompt caching, the developer tested Kimi 2.6, DeepSeek 4 Pro, and DeepSeek 4 Flash before settling on Qwen as the best fit for handling the multi-hundred-thousand-line codebase.

Qwen 3.6 35B-A3B accepts image input and screenshots, a requirement for the developer's workflow. While the hardware to run the 35-billion-parameter model locally exists, OpenRouter's hosted inference avoided upfront infrastructure costs. The effective token price reflects OpenRouter's prompt caching, which reduces repeat charges for the large context window the codebase requires. The model's instruction-following and multimodal capabilities proved more reliable than the alternatives tested, though the developer noted minor issues with each competitor that Qwen avoided.

Two gaps remain: Cursor's cloud agent orchestration and the high-throughput auto-complete that Composer 2 provides. Those features are the main reason to stay on a proprietary IDE, but the cost and capability trade-off now favors the open-weight model for this use case. The developer's report adds to growing anecdotal evidence that Qwen 3.6 35B-A3B punches above its parameter count on real-world coding tasks. The next test will be whether Qwen's upcoming releases close the agent orchestration gap or whether third-party tooling around the model catches up to what Cursor ships out of the box.