GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
cvis-tmu
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
cvis-tmu
Model Tree
Qwen/Qwen2.5-VL-7B-Instruct
Input Modalities
Output Modalities
Supported Functionality
cc594898137f460bfe9f0759e9844b3ce807cfb5.grillium), node trig0033, 4 H100 GPUs.1019445; wall time: 20h 36m 27s.| Step | Validation loss |
|---|---|
| 200 | 1.010405 |
| 400 | 0.971601 |
| 600 | 0.961912 |
| 617 | 0.961902 |
Mean training loss: 1.037220. These metrics were measured with the base model plus LoRA adapter before merging. The merged export has not been separately benchmarked; BF16 merging can introduce small numerical differences. No downstream task accuracy is claimed.
Use this repository as model_name_or_path with a Qwen2.5-VL-compatible Transformers runtime. For LLaMA-Factory inference, use template: qwen2_vl and omit adapter_name_or_path. The model, processor, tokenizer, and chat template are packaged together.
Exported offline through LLaMA-Factory llamafactory-cli export with infer_dtype=bfloat16, export_device=cpu, and safetensors shards of up to 5 GB. This repository contains inference weights, not optimizer state; preserve the original adapter checkpoint to resume the original training run.
The base model is licensed under Apache 2.0. Dataset terms continue to apply to their respective data. This is a research fine-tune; assess performance and suitability for the intended use.