GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
zhongweixie
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Run this model inference with full control and performance in your environment.
Model Details
Model Provider
zhongweixie
Model Tree
Qwen/Qwen3-VL-8B-Instruct
Input Modalities
Output Modalities
Supported Functionality
| 项 | 值 |
|---|---|
| 实验名 | exp_rl_v5_v13_gdpo |
| 算法 | GDPO |
| 变体 | 冷启动基线。无 rubric overlay，无 dim scoring。 |
| 起点 | 原始 Qwen3-VL-8B-Instruct(冷启动，非 SFT checkpoint) |
| 最终 checkpoint | iter_0000560 |
| 训练量 | 561 个 rollout step，约 3 个 epoch，187 条 prompt |
| samples / prompt | 8 |
| rollout batch | 1 |
| global batch | 8 |
| 硬件 | 8xH800，TP=8，colocate |
| 奖励 judge | qwen/qwen-2.5-72b-instruct (OpenRouter) |
学习率、KL 系数等由 slime 的 formal profile 提供，未在启动脚本中显式设置，
故此处不列出，以免给出未经核实的数值。
与同期 5 个 *_stage3 模型的区别：那些以 stage3 SFT checkpoint
(exp_stage3_v2_merged) 为起点，本模型是从原始 instruct 冷启动的。
目前没有可信的 held-out 评测数字。 不要把本模型当作已验证优于其起点。
原因：
ninja 未能启动，缺少可比基线。python
from transformers import Qwen3VLForConditionalGeneration, AutoProcessormodel = Qwen3VLForConditionalGeneration.from_pretrained("zhongweixie/qwen3vl-8b-claw-rl-gdpo-v13-iter560", dtype="auto", device_map="auto")processor = AutoProcessor.from_pretrained("zhongweixie/qwen3vl-8b-claw-rl-gdpo-v13-iter560")
由 slime/tools/convert_torch_dist_to_hf.py 从 Megatron 分布式 checkpoint
/aifs4su/hansirui_3rd/rl-checkpoints/exp_rl_v5_v13_gdpo/iter_0000560
转换而来，缺失权重与 tokenizer/config 取自原始 Qwen3-VL-8B-Instruct。