maveryn/trace-qwen2.5-vl-7b

VISIONPricing:Input $0.4 / Output $0.8Concurrent Unit Cost:1Model Size:7BQuant:FP8Context Size:32kPublished:Jul 16, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

maveryn/trace-qwen2.5-vl-7b is a 7 billion parameter vision-language model developed by maveryn, derived from Qwen2.5-VL-7B-Instruct. It is a GRPO (Grounded Reasoning Policy Optimization) checkpoint, fine-tuned on 64,000 grounded visual-reasoning examples across 1,000 tasks from the maveryn/trace dataset. This model specializes in improving visual reasoning capabilities and structured JSON output for multimodal tasks, demonstrating a significant performance uplift on the trace_eval_v1 benchmark.

Loading preview...

TRACE Qwen2.5-VL 7B Overview

TRACE Qwen2.5-VL 7B is a 7 billion parameter vision-language model developed by maveryn, built upon the Qwen/Qwen2.5-VL-7B-Instruct base model. It leverages Grounded Reasoning Policy Optimization (GRPO) and was fine-tuned on 64,000 grounded visual-reasoning examples from the maveryn/trace dataset, spanning 1,000 distinct tasks. This model is specifically designed to enhance multimodal reasoning and generate structured JSON outputs for visual queries.

Key Capabilities

  • Enhanced Visual Reasoning: Significantly improves performance on complex visual reasoning tasks compared to its base model.
  • Structured Output: Optimized to produce answers in a canonical JSON format, specifically {"answer": ...}.
  • Grounded Responses: Trained on a dataset focused on grounded visual reasoning, ensuring responses are tied to visual evidence.
  • Benchmark Improvement: Achieved an overall score of 51.99 ± 0.17 on the trace_eval_v1 benchmark, representing a +4.06 ± 0.41 paired improvement over the base Qwen2.5-VL-7B-Instruct model.

Good For

  • Research on Multimodal Reasoning: Ideal for academic and research purposes exploring advanced visual-language understanding.
  • Verifiable-Reward Post-Training: Suitable for experiments in reinforcement learning from human feedback (RLHF) or similar verifiable reward mechanisms.
  • Applications Requiring Structured Visual Answers: Useful for scenarios where precise, machine-readable JSON outputs are needed for visual queries.
  • Benchmarking Visual-Language Models: Can serve as a strong baseline or comparison model for evaluating new visual reasoning techniques.

Finetunes
1 models