Qwen 3.6-35B-A3B hits 140 t/s on RTX 4090

Qwen 3.6-35B-A3B is a category-defining "agentic first" open model that brings state-of-the-art coding performance to consumer hardware.

• –The Mixture-of-Experts (MoE) design uses only 3B active parameters, allowing it to run at high speed even with high-precision Q8 quantization.
• –Native 262k context window and "thinking preservation" feature reduce redundant computation in long-running agentic tasks.
• –It excels at repository-level reasoning and tool calling, directly challenging proprietary models like Claude 3.5 Sonnet for local workflows.
• –Multimodal support allows the model to reason about UI/UX designs and diagrams alongside code.