GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
ngqtrung
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
ngqtrung
Model Tree
Qwen/Qwen3.5-9B
Input Modalities
Output Modalities
Supported Functionality
Checkpoints every 100 steps were each evaluated (pass@1, T 0.6, top-p 0.95, top-k 20, 32K response cap, robust answer matching). Step 500 has the best mean score of this run.
| benchmark | Qwen3.5-9B | this model |
|---|---|---|
| videomme | 0.628 | 0.670 |
| longvideobench | 0.566 | 0.607 |
| video_holmes | 0.475 | 0.488 |
| tempcompass | 0.752 | 0.745 |
| vsibench | 0.480 | 0.478 |
| mean | 0.605 | 0.627 |
Scores: accuracy with a robust answer matcher (official score where the matcher does not apply). On the image domains
SFT stays below the post-trained base; it shortens the answers (base 1.2K-15K tokens, SFT ~130-840) and fixes the format
(</think> closed and \boxed{} in > 99 % of answers). The next stage (GSPO RL from this checkpoint) is meant to
recover and exceed the base.
python
from transformers import AutoProcessor, AutoModelForImageTextToTextmodel = AutoModelForImageTextToText.from_pretrained("ngqtrung/Qwen3.5-9B-MOPD-SFT-vid", torch_dtype="bfloat16", device_map="auto")processor = AutoProcessor.from_pretrained("ngqtrung/Qwen3.5-9B-MOPD-SFT-vid")# processor.apply_chat_template(messages, enable_thinking=True, add_generation_prompt=True, ...)vLLM: vllm serve ngqtrung/Qwen3.5-9B-MOPD-SFT-vid (stop token ids 248046, 248044 from generation_config.json).
Part of the collection MOPD: multi-teacher on-policy distillation (Qwen3.5-9B).