Will LongWriter-Zero Be Open-Sourced, and Can It Be Trained with Limited GPU Resources?

[PirateX0]

Hi, thanks for the great work on LongWriter-Zero—this is a very exciting project.

I have a couple of questions regarding both the open-source plan and the training setup under limited GPU resources.

1. Open-source timeline

Is there a plan to open-source the LongWriter-Zero implementation?
If so, could you share an approximate timeline (e.g., in the coming weeks or months, or after a certain milestone)? This would be very helpful for planning reproduction and follow-up experiments.

2. Training with smaller models and fewer GPUs

I’m interested in experimenting with LongWriter-Zero under more constrained hardware settings. Specifically, would it be feasible to apply the method to a smaller model such as Qwen3-4B, instead of Qwen2.5-32B, to reduce GPU requirements?

For reference, I currently have access to 4×80GB A800 GPUs, which is significantly less than the 8 nodes × 8 H800 GPUs setup described in the paper.

In this context, I’m wondering:

• Would a 4B-scale model still meaningfully benefit from the LongWriter-Zero RL approach?
• Are there any recommended adjustments for smaller models and limited hardware, such as:
  • number of concurrent trajectories per optimization step
  • batch size
  • maximum output length
  • or other key hyperparameters that are important for stable RL training?

Any guidance or high-level suggestions would be greatly appreciated.
Looking forward to the open-source release—thanks again for the impressive work!