Qwen 3.6 Q4 hits 112 t/s in Pi harness

The sparse MoE architecture in Qwen 3.6-35B-A3B is a game-changer for local agentic workflows, proving that Q4 quantization is now the professional-grade sweet spot for speed and context stability.

• –Q4 quantization offers a massive 2.2x speedup (112 t/s vs 50 t/s) with negligible impact on coding logic, allowing for near-instant agent feedback loops.
• –The model maintains remarkable consistency through multiple "compacting" cycles at 131k context, solving the long-context retrieval issues that plagued earlier 2.5-series quants.
• –Using the minimalist Pi harness reduces system prompt overhead to under 1,000 tokens, maximizing the available window for repository-level reasoning.
• –Sparse MoE (3B active parameters) allows for high-throughput inference on consumer GPUs while retaining the logic depth of a much larger dense model.
• –For most coding tasks, the "perplexity tax" of Q4 is offset by the ability to run more agentic verification loops per minute compared to the slower Q8.