GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
hyperquader
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
hyperquader
Model Tree
mistralai/Ministral-3-3B-Instruct-2512-BF16
Input Modalities
Output Modalities
Supported Functionality
The adapter alone, for anyone who wants to inspect it, merge it themselves or train on from it. What it was trained on, how and what it scores are on the merged model's card; nothing is restated here.
| file | what it is |
|---|---|
adapter_config.json, adapter_model.safetensors | the LoRA, r 16, alpha 32, the first epoch |
tokenizer.json, tokenizer_config.json | the base's tokenizer, unchanged |
chat_template.jinja | the template every training target was rendered with; calls do not parse without it |
The targets are named by the base's module paths, which put the language
layers under model.language_model.layers.N. The vision tower reuses the
same projection names and is not adapted.
python
from peft import PeftModelfrom transformers import AutoModelForImageTextToText base = AutoModelForImageTextToText.from_pretrained( "mistralai/Ministral-3-3B-Instruct-2512-BF16", dtype="bfloat16")model = PeftModel.from_pretrained( base, "hyperquader/ministral-3-3b-amberlin-tools-adapter").merge_and_unload()Apache 2.0, inherited from the base model.