GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
C1oudX
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
C1oudX
Model Tree
Qwen/Qwen3.5-9B
Input Modalities
Output Modalities
Supported Functionality
c202236235762e1c871ad0ccb60c8ee5ba337b9a.OR results exclude exact question-text matches with both train and validation, leaving 1393 tasks. OPT-Bench ML uses fixed local holdouts; OPT-Bench NP evaluates constructive feasibility and objectives.
OR 结果排除与 train/val 题目文本精确重叠的样本，共 1393 题。ML 和 NP 任务分别来自 OPT-Bench 的 ML 与 NP 部分。
| Measure | Untrained base | This step36 model |
|---|---|---|
| OR1393 micro accuracy | 777/1393 (55.78%) | 961/1393 (68.99%) |
| OR1393 macro accuracy over eight sets | 44.55% | 62.36% |
| OPT-Bench ML valid local-holdout submissions | 16/20 | 18/20 |
| OPT-Bench NP feasible submissions | 24/50 (48%) | 30/50 (60%) |
Both models use eight tools, greedy n=1, up to 15 assistant turns and 8192-token prompt/response caps, including tool observations in the response budget. Among 14 jointly valid OPT-Bench ML tasks, step36 wins 3 metrics and loses 11. Among 22 jointly feasible OPT-Bench NP cases, it wins 1 objective, loses 1 and ties 20.
Per-task results are in evaluation_results.json.
Use the AGRPO environment locks and tool adapters to reproduce agentic evaluation. The recorded runtime uses Transformers 5.12.1, PyTorch 2.10.0 and vLLM 0.19.0.
python
from transformers import AutoModelForImageTextToText, AutoProcessor model_id = "C1oudX/AGRPO-Qwen3.5-9B-step36"processor = AutoProcessor.from_pretrained(model_id)model = AutoModelForImageTextToText.from_pretrained( model_id, dtype="auto", device_map="auto")Apache-2.0. The base revision's LICENSE is included. Training metadata is in training_metadata.json.