AdaVaR-3B Beats GPT-4o on Visual Reasoning and Runs on Local Hardware
A quantized GGUF build of AdaVaR-3B lets you run a Qwen2.5-VL fine-tune that picks between text reasoning and image-grounded reasoning on the fly.
- GGUF quants of AdaVaR-3B, a Qwen2.5-VL fine-tune with adaptive text or grounded reasoning modes.
- Picks between natural-language chain of thought and image coordinate grounding per question.
- AdaVaR-7B hits 55.82% average over 8 benchmarks, beating GPT-4o at 53.20%.
- AdaVaR-3B matches Qwen2.5-VL-7B while using less than half the parameters.
- Trained via supervised cold start plus AdaGRPO reinforcement learning with mode-relative advantage.
- Apache 2.0, runs under llama.cpp or LocalAI with F16 vision projector.
A third-party maintainer has published GGUF quantizations of the original AdaVaR-3B model. The new GGUF repository packages the Qwen2.5-VL-based checkpoint for llama.cpp-compatible runtimes, allowing the model to choose between text reasoning and image-grounded coordinate reasoning on local hardware.
Quantization stores weights at lower precision, reducing disk and memory requirements while risking some quality loss. The release changes deployment format and numerical precision while preserving AdaVaR’s architecture and training. Its main collection spans Q2-class quants through F16, with more aggressive IQ variants available separately.
- Released model: AdaVaR-3B, based on Qwen2.5-VL-3B
- Local format: GGUF model files with a separate vision projector
- Runtime target: Recent llama.cpp-family tools with Qwen2.5-VL support
- License: The model and quant repositories list Apache-2.0
A router between words and boxes
AdaVaR can produce two kinds of reasoning traces. Text mode expresses intermediate steps in natural language. Grounded mode emits image coordinates, such as bounding boxes, that connect parts of the reasoning to specific regions. The model outputs coordinate references rather than modifying or drawing on the source image.
For each image and question, a learned router first selects a mode and then conditions answer generation on that choice. The authors also describe an inference fallback that detects repetitive reasoning without a final answer, switches modes, and retries.
The architecture is described in the Mixture-of-Visual-Thoughts paper, submitted to ICLR 2026. Two research checkpoints were trained: AdaVaR-3B on Qwen2.5-VL-3B and AdaVaR-7B on Qwen2.5-VL-7B. The current GGUF release covers the 3B checkpoint.
Teaching the router to choose
Training begins with supervised examples that teach both reasoning modes. A reinforcement-learning stage then applies the authors’ AdaGRPO method to improve mode selection:
- Prefix-guided exploration generates multiple rollouts in each mode for every training question, giving the router evidence about both paths.
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
