Sparsified Qwen 3.6-35B A3B models maintain performance parity with the baseline model on coding tasks across context lengths up to 32,000 tokens