havok2/Qwen-AgentWorld-35B-A3B-VL36

TEXT GENERATIONConcurrent Unit Cost:3Model Size:35.1BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jun 26, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

havok2/Qwen-AgentWorld-35B-A3B-VL36 is a 35.1 billion parameter multimodal variant created by havok2, combining the agentic text backbone of Qwen-AgentWorld-35B-A3B with the vision tower from Qwen3.6-35B-A3B. This model, built through weight-grafting without additional training, integrates vision capabilities into an existing agentic text model. It is designed for applications requiring both advanced text-based agency and image/video input processing, leveraging a 32768-token context length.

Loading preview...

Overview

havok2/Qwen-AgentWorld-35B-A3B-VL36 is a 35.1 billion parameter multimodal model created by havok2. It is a weight-grafted variant, combining the agentic text backbone from Qwen/Qwen-AgentWorld-35B-A3B with the vision tower (ViT + projector) from Qwen/Qwen3.6-35B-A3B. This model was created through "weight surgery" without any additional training, preserving the original agentic and tool-calling capabilities while adding vision.

Key Capabilities

  • Agentic Text & Tool-Calling: Retains the full agentic and tool-calling skills of the Qwen-AgentWorld backbone, as its text weights remain untouched.
  • Image / Video Input: Supports multimodal input, including images and videos, with basic smoke tests showing accurate grounding for shapes, colors, and OCR.
  • Full Precision BF16: The model operates in full precision BF16.

Limitations and Considerations

  • Experimental Cross-Generation Graft: The vision tower is from Qwen3.6, while the AgentWorld backbone is a Qwen3.5 derivative. While dimensionally compatible, it is not a natively co-trained VLM.
  • Vision Quality Not Benchmarked: The vision capabilities have not been formally benchmarked, and performance may differ from natively trained VLMs.
  • No Additional Safety Tuning: The model has not undergone additional safety tuning or evaluation beyond basic smoke tests.

Usage

This model can be served using vLLM (version 0.23 or newer) and requires approximately 70 GB of VRAM in BF16, fitting on 4x24 GB GPUs with --tensor-parallel-size 4.

Finetunes
1 models