Qwen 3.6 35B Genesis-V2 hits with APEX, MTP

This release demonstrates that open-weight models often require community "repair" to reach their full potential after being damaged by training or conversion drifts.

• –APEX quantization uses Wasserstein distance metrics to restore weight symmetry and fix saturation issues in the base model.
• –Native MTP support significantly increases throughput on compatible hardware by predicting multiple tokens in a single forward pass.
• –The 35B MoE architecture with 3B active parameters remains the "goldilocks" size for high-performance local AI on consumer GPUs.
• –Tester results indicate high reliability across the 262K context window, making it a viable candidate for complex long-context coding tasks.
• –Hybrid Gated DeltaNet and Softmax attention provides a scalable alternative to traditional quadratic Transformer overhead.