GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
kiel2
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
kiel2
Model Tree
Input Modalities
Output Modalities
Supported Functionality
KielLens-VQA is a high-performance, fine-tuned vision-language model designed for precise visual question answering, spatial reasoning, and image-grounded analysis. Built upon the state-of-the-art Qwen2.5-VL-7B-Instruct architecture, KielLens-VQA excels at interpreting complex visual data, extracting key insights, and providing detailed context-aware answers to user queries.
You can load the model and its adapters seamlessly using transformers and peft:
python
import torchfrom transformers import AutoModelForVision2Seq, AutoProcessorfrom peft import PeftModel base_model_id = "unsloth/Qwen2.5-VL-7B-Instruct" # Load base vision-language modelbase_model = AutoModelForVision2Seq.from_pretrained( base_model_id, torch_dtype=torch.bfloat16, device_map="auto") # Load KielLens-VQA adaptersmodel = PeftModel.from_pretrained(base_model, "kiel2/KielLens-VQA")processor = AutoProcessor.from_pretrained("kiel2/KielLens-VQA") Training ProcedureThe model was fine-tuned using the Unsloth framework and TRL library with the following configurations: Quantization: 4-bit (via NF4) for memory efficiency and high-speed inference. Target Modules: Vision and attention layers optimised for multimodal token mapping. Optimization: Paged AdamW 32-bit. Precision: Mixed precision (bf16). Limitations and Ethical ConsiderationsInput Sensitivity: Performance may vary depending on image resolution, lighting quality, and framing. Generalization: While optimized for robust visual reasoning, the model should be deployed with human oversight for mission-critical visual analytics tasks.import torchfrom transformers import AutoModelForVision2Seq, AutoProcessorfrom peft import PeftModel base_model_id = "unsloth/Qwen2.5-VL-7B-Instruct" # Load base vision-language modelbase_model = AutoModelForVision2Seq.from_pretrained( base_model_id, torch_dtype=torch.bfloat16, device_map="auto") # Load KielLens-VQA adaptersmodel = PeftModel.from_pretrained(base_model, "kiel2/KielLens-VQA")processor = AutoProcessor.from_pretrained("kiel2/KielLens-VQA") Training ProcedureThe model was fine-tuned using the Unsloth framework and TRL library with the following configurations: Quantization: 4-bit (via NF4) for memory efficiency and high-speed inference. Target Modules: Vision and attention layers optimised for multimodal token mapping. Optimization: Paged AdamW 32-bit. Precision: Mixed precision (bf16). Limitations and Ethical ConsiderationsInput Sensitivity: Performance may vary depending on image resolution, lighting quality, and framing. Generalization: While optimized for robust visual reasoning, the model should be deployed with human oversight for mission-critical visual analytics tasks.