cagataydev/strands-qwen3-vl-2b

VISIONConcurrent Unit Cost:1Model Size:2BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jul 14, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

The cagataydev/strands-qwen3-vl-2b is a 2 billion parameter Qwen3-VL-2B model, post-tuned by cagataydev, specifically designed for the Strands Agents framework. This model is optimized for continuous learning after deployment, featuring surprise-gated weight updates and EMA-bounded plasticity. It excels as a domain expert for Strands Agents, demonstrating improved performance on domain-specific tasks and the ability to learn new facts without forgetting prior knowledge.

Loading preview...

Model Overview

The cagataydev/strands-qwen3-vl-2b is a 2 billion parameter Qwen3-VL-2B model, post-tuned by cagataydev, specifically for the Strands Agents ecosystem. It has been fine-tuned using LoRA on the entire public codebase and documentation of strands-agents and strands-labs, encompassing approximately 5.6 million tokens. This model's core innovation lies in its integration with the strands-slm plastic self-learning layer, enabling it to continually learn and adapt after deployment.

Key Capabilities

  • Continuous Learning: The model can update its weights at inference time through surprise-gated mechanisms, allowing it to learn from new data streams.
  • Domain Expertise: Demonstrates significant improvement as a domain expert for Strands Agents, with reduced negative log-likelihood (NLL) on relevant probes and corpus holdouts.
  • Forgetting Prevention: Maintains its acquired expertise even after learning from out-of-distribution (OOD) data streams.
  • Provable Off-Switch: Features a reset() function that returns the model to its base state, ensuring bit-identical reproducibility.
  • Agent Competence Growth: Shows improved performance on held-out tasks for Strands Agents after curated lessons.

Good For

  • Developers building Strands Agents who require a model that can adapt and learn in real-time.
  • Applications needing a vision-language model with specialized knowledge in the Strands Agents framework.
  • Use cases where continual learning and plasticity are critical for model performance and evolution.