qwen3-8b-dmax

DMax/OPUT-trained variant of Qwen/Qwen3-8B.

• Base: Qwen/Qwen3-8B
• Training framework: dllm-jax (JAX/Flax NNX, TPU)
• Source checkpoint: gs://dllm-jax-europe-west4/checkpoints/qwen3-8b-PRODUCTION-mix-dmax-opUT-optionB-v6e64-20260504-0500/checkpoint_1000
• Step: 1000

This is a JAX-trained DMax/OPUT (block-diffusion / on-policy under-tuning) fine-tune. The weights are stored in standard HF Qwen3ForCausalLM layout — the same loader path as the base model. Inference requires the dllm-jax DMax block-diffusion path (the model is not an autoregressive Qwen3 — it expects a doubled [noised; clean] input under a block-diffusion mask).