zwc2003/DriveMA-2B

VISIONPricing:Input $0.32 / Cached $0.064 / Output $1.6Concurrent Unit Cost:1Model Size:2.3BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jul 18, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

DriveMA-2B is a 2.3 billion parameter vision-language-action model developed by zwc2003, fine-tuned from Qwen3.5-2B. It is specifically designed for autonomous driving planning, formulating it as a two-turn generation process involving meta-actions and waypoints. This model excels at driving planning tasks by predicting interpretable meta-actions from multi-view observations and vehicle state. It is intended for research in vision-language-action modeling and end-to-end autonomous driving planning.

Loading preview...

DriveMA-2B: Vision-Language-Action Model for Autonomous Driving

DriveMA-2B is a 2.3 billion parameter model developed by zwc2003, fine-tuned from Qwen3.5-2B. It is the official checkpoint for the research detailed in the paper "DriveMA: Driving Vision-Language-Action Models with Verifiable Meta-Actions." The model is specifically engineered for autonomous driving planning through a unique three-stage fine-tuning pipeline: action-centric pretraining, action-conditioned trajectory supervised fine-tuning, and turn-level reinforcement learning.

Key Capabilities

  • Two-Turn Driving Planning: Formulates driving planning as a two-turn generation process, first predicting a compact, interpretable meta-action from multi-view observations and vehicle state, then generating future waypoints conditioned on that meta-action.
  • Vision-Language-Action Integration: Combines visual inputs, language understanding (for meta-actions), and action generation for complex driving scenarios.
  • Performance: Achieves competitive results on the Waymo Open Dataset vision-based end-to-end planning benchmark, with reported RFS Overall of 8.060 and ADE@5s of 2.616.

Intended Use Cases

  • Research: Ideal for academic and research purposes in vision-language-action modeling and end-to-end autonomous driving planning.
  • Development: Can be used as a base for developing and experimenting with novel autonomous driving algorithms, particularly those involving meta-actions and verifiable planning.

Important Considerations

This model is not validated for deployment in safety-critical systems and should not be used to control a real vehicle without independent safety validation and compliance with regulations. Users must obtain source image/video assets for datasets under their original licenses.

Finetunes
1 models