Open-dLLM tests Qwen 3.6 diffusion

This is less a launch than a serious notebook from the edge of what single-card hardware can support. If the alignment and training tricks hold up, diffusion LLMs stop being a lab curiosity and start looking like an actual local-model acceleration path.

• –The key signal is hardware feasibility: the post is trying to shrink an apparently massive memory footprint into something a 5090 can handle.
• –Open-dLLM’s representation alignment approach and the d3LLM-inspired MDM loss are two different attempts to cut denoising steps without losing too much quality.
• –The claimed 4x speedup on Qwen 2.5 matters, but the bigger question is whether that transfers cleanly to Qwen3.6-27B under the same “same weights, new decoding” framing.
• –If the helper scripts for anchors and trajectories are solid, this could become a useful experimental stack for other researchers chasing faster diffusion decoding.
• –The cable-burning PSU warning is not a joke; this is very much frontier tinkering on consumer hardware, not a turnkey product.