πŸ“„ arXiv Β· πŸ’» GitHub

Latest checkpoint

Updated 2026-10-28 to online step 250, continuing from Stage II step 14925. Only the REMORY memory network is updated; Qwen3.8-27B remains frozen. The architecture and inference interface are unchanged. The release contains FP32 Safetensors weights, with lossless tensor export and CPU forward parity verified against the native serving checkpoint. Details: release.json.

This online checkpoint has not been re-evaluated on the paper benchmarks; the previously reported results should not be read as measurements of this revision. The preceding Stage II checkpoint remains available at f24de2ac3996da36678cd5d196a7dc0dbbc3aa34.

Citation

@misc{xia2026remory,
  title = {Remory: Learning Residual Memory for Context Compaction},
  author = {Hanchen Xia and Baoyou Chen and Yutang Ge and Naihao Deng and
            Senqiao Yang and Zilong Dong and Weihao Yuan and Siyu Zhu},
  year = {2026},
  eprint = {2610.11287},
  archivePrefix = {arXiv},
  primaryClass = {cs.CL},
  url = {https://arxiv.org/abs/2610.11287}
}
Downloads last month
3,257
Safetensors
Model size
2B params
Tensor type
F32
Β·
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for fudan-generative-ai/Qwen3.8-27B-REMORY-1.9B

Base model

Qwen/Qwen3.8-27B
Adapter
(275)
this model

Collection including fudan-generative-ai/Qwen3.8-27B-REMORY-1.9B

Paper for fudan-generative-ai/Qwen3.8-27B-REMORY-1.9B