GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
GL3MON
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
GL3MON
Model Tree
GL3MON/qwen3-vl-2b-medical-pmcvqa-sft-merged
Input Modalities
Output Modalities
Supported Functionality
Evaluated on all 2,000 OctoMed/PMC-VQA test examples with two attempts per question.
| Model | Pass@1 | Pass@2 | Failed examples |
|---|---|---|---|
| Saved base run | 30.40% | 41.05% | 5 |
| Saved SFT run | 41.55% | 56.00% | 5 |
| GRPO checkpoint-2100 | 41.35% | 56.85% | 0 |
The checkpoint produced 4,000 retained successful inference prompts, using 1,235,914 input tokens and 1,373,739 output tokens, with 0 generation truncations.
Pass@1 checks the first answer. Pass@2 checks whether either attempt matches the ground-truth option; it is an oracle-style metric and does not provide a method for selecting the correct answer at inference time. Failed examples count as incorrect.
The two attempt temperatures were 0.6 and 0.75, with top-p 0.8, top-k 20, thinking enabled, and at most 4,096 completion tokens per attempt. Images were sent as JPEG at quality 95. The server used bfloat16, an 8,192-token context, Triton language attention, PyTorch vision attention, and 16 concurrent benchmark workers.
These are historical comparisons from stochastic runs. Input-token totals and processor/server configurations differ, so the small accuracy differences do not establish a controlled improvement from GRPO.
See evaluation summary, combined metrics, and adapter-merge checks.
python
import torchfrom transformers import AutoProcessor, Qwen3VLForConditionalGeneration repo = "GL3MON/qwen3-vl-2b-medical-pmcvqa-grpo-merged"processor = AutoProcessor.from_pretrained(repo)model = Qwen3VLForConditionalGeneration.from_pretrained( repo, dtype=torch.bfloat16, device_map="auto")Use the processor's chat template for image-and-text inputs. The model and processor were exported with Transformers 5.6.2. The merged checkpoint was validated for finite representative vision and language weights and for updates from the SFT base.