GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
woooooz
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
woooooz
Model Tree
Qwen/Qwen-Drive-1.0-4B
Input Modalities
Output Modalities
Supported Functionality
This repository contains the driving-adapted Qwen3.5 VLM, without the external BEV perception head or SFT/RL planning experts. It does not provide autonomous-driving trajectory outputs.
The conversion extracts vlm_config as the top-level configuration, preserving the source Transformers version metadata, and removes exactly the vlm. prefix from every weight name. The safetensors tensor payload, shapes, dtypes, and offsets remain unchanged. Tokenizer, vocabulary, image preprocessor and chat templates are copied unchanged. No additional training, fine-tuning, merging or quantization was performed.
Source revision: 28484089a7cc8c335cf5089fb0745cf7c49b6eaa.
(1, 2560) outputs in a local retrieval wrapper, using last-token pooling and L2 normalization.See verification.json for hashes and provenance. These checks establish packaging and loading correctness; they do not establish retrieval quality, training stability or driving safety. This checkpoint has no dedicated retrieval embedding head or contrastive embedding training added by this conversion.
Tested with Transformers 5.6.0.dev0 and PyTorch 2.6.0+cu124 on NVIDIA A100. Use a Transformers release that supports Qwen3.5.
python
import torchfrom transformers import AutoProcessor, Qwen3_5ForConditionalGeneration repo = "woooooz/Qwen-Drive-1.0-4B-VLM"processor = AutoProcessor.from_pretrained(repo)model = Qwen3_5ForConditionalGeneration.from_pretrained( repo, dtype=torch.bfloat16, attn_implementation="sdpa",).to("cuda").eval() # For hidden-state extraction, model.model is the Qwen3_5Model backbone.# Feed processor outputs to model.model(..., return_dict=True, use_cache=False)# and obtain outputs.last_hidden_state.The original weights tie the language-model output projection to input embeddings, so the safetensors file does not store a separate lm_head.weight. The standard Qwen3.5 loading path restores the tied relationship.
Apache-2.0, inherited from the original Qwen-Drive release. See LICENSE. All original model credit belongs to the Qwen-Drive authors; this repository only repackages the VLM component.