GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
Karl28
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
Karl28
Model Tree
Qwen/Qwen2.5-7B-Instruct
Input Modalities
Output Modalities
Supported Functionality
The base model collected 8 trajectories on each of 500 training tasks. GLM-5.2 wrote an episode summary and reusable
skill for every trajectory, and the model was fine-tuned to produce these analyses for 3 epochs, with learning rate
5e-6 and global batch size 8.
The train and validation sets contain 3,587 and 399 examples. Full trajectories and trajectory-to-summary/skill annotations are included in the dataset repository.
The safetensors in this repository are the original initialization checkpoint files used by our experiments. Their
tensors are FP32, while config.json retains torch_dtype: bfloat16 from the export configuration. We keep this
combination unchanged for exact reproducibility. The RSIPO launchers use bfloat16 rollout/inference settings.
bash
hf download Karl28/RSIPO-SFT-Qwen2.5-7B-Instruct-Search --local-dir $MODELS_ROOT/RSIPO-SFT-Qwen2.5-7B-Instruct-Searchbash scripts/rsipo/run_rsipo.sh search RSIPO-SFT-Qwen2.5-7B-Instruct-SearchFor search-based QA, also set RSIPO_SEARCH_TRAIN_DATA and start the retriever as described in the code README.
Apache License 2.0 (LICENSE), the license of the base model Qwen/Qwen2.5-7B-Instruct. The weights are modified from the base model by fine-tuning.
bibtex
@misc{zhang2026rsipo, title={RSIPO: Recursive Self-Improving Policy Optimization for Agentic Reinforcement Learning via Skill-Policy Co-Evolution}, author={Fan Zhang and Hao Wu and Jinyang Wu and Hongyu Qu and Zheng Lian and Xiao Luo and Yefeng Zheng and Bo Zhang and Haoxuan Li and Pheng-Ann Heng}, year={2026}, eprint={2610.15772}, archivePrefix={arXiv}, primaryClass={cs.CL}, url={https://arxiv.org/abs/2610.15772},}