R2Orch-9B

Paper: arXiv:2610.22740 · Hugging Face Paper Page

Code: beita6969/R2Orch

The trained R²Orch Director as a complete model. The Director's LoRA adapter (rank 64, alpha 64, on q_proj, k_proj, v_proj, o_proj of the full-attention layers and on gate_proj, up_proj, down_proj of every MLP) is merged into Qwen3.5-9B. The Workers of R²Orch are the unmodified Qwen3.5-9B; the Proposer is a training-time component and is not part of this release. The 128 merged weight matrices are stored in float32; every other tensor is the unchanged bfloat16 base weight. The LoRA update is about 0.13% of the weight norm (mean over the merged matrices, maximum 0.19%), below the roughly 0.4% relative resolution of bfloat16, so load the model in float32 to reproduce the trained Director exactly.

import torch
from transformers import AutoTokenizer, Qwen3_5ForConditionalGeneration
model = Qwen3_5ForConditionalGeneration.from_pretrained("beita6969/R2Orch-9b", dtype=torch.float32)
tokenizer = AutoTokenizer.from_pretrained("beita6969/R2Orch-9b")

The merge was checked against the unmerged base + adapter in float32: the next-token logits differ by at most 0.004 and the KL divergence is below 1e-7.