OmniThoughtVis: A Large-Scale Distillation Framework for Deployable Multimodal Reasoning Models

OmniThoughtVis is an extensible data curation and knowledge distillation pipeline. Its core goal is to bridge the gap between large models (which have strong reasoning capabilities but are hard to deploy) and small models (which are easy to deploy but lack high-quality multimodal chain-of-thought data). Through structured chain-of-thought generation, difficulty-aware selection, and label diversity sampling, this framework constructs a high-quality multimodal reasoning dataset with 1.8 million samples, successfully transferring the reasoning capabilities of large models to small models with 2B-8B parameters, and providing a feasible path for the practical deployment of multimodal reasoning models.

In recent years, multimodal large language models (MLLMs) have demonstrated strong chain-of-thought (CoT) capabilities in vision-language reasoning tasks, but their high computational cost and inference latency make them difficult to directly deploy in production environments. Smaller MLLMs infer faster, cost less, and are easy to deploy on edge devices, but they lack large-scale, high-quality multimodal chain-of-thought supervision data—multimodal reasoning data annotation is complex and expensive, making it hard to obtain through manual annotation, thus forming a deployment gap between large and small models.

OmniThoughtVis's large-scale distillation pipeline includes three key stages: seed pool construction (diverse open-source data), structured chain-of-thought generation (teacher model generates trajectories with reasoning steps), and joint annotation (reasoning difficulty, answer quality, semantic task labels). To ensure large-scale data quality, the framework designs a three-layer filtering mechanism: rule-based filtering (quickly eliminating low-quality samples), difficulty-aware selection (maintaining a balance of easy/medium/hard samples), and label diversity sampling (covering various tasks and scenarios). Finally, it constructs a 1.8 million high-quality dataset, supporting the extraction of controllable subsets as needed.

The research team used OmniThoughtVis to distill and train the Qwen3-VL series models (2B-8B parameters), and the evaluation results on nine multimodal reasoning benchmarks are impressive: the 4B model improved by +16.8 points on the MathVerse benchmark and +5.6 points on the MMMU-Pro benchmark; distillation gains exist across different parameter scales, proving universality; more notably, the distilled 4B model reached or even exceeded the undistilled 8B baseline model in multiple tasks, changing the trade-off between model scale and capability.

The success of OmniThoughtVis comes from three major technical insights: 1. Structured chain-of-thought provides process supervision, allowing student models to learn reasoning processes rather than just answers; 2. Data quality is better than quantity—carefully selected high-quality samples are more effective than a large number of low-quality samples; 3. Controllable data construction supports flexible extraction of subsets by difficulty and task type, facilitating optimized training for specific scenarios.

OmniThoughtVis has significant value for practical deployment: 1. Cost-effectiveness optimization—replacing an 8B model with a 4B model can significantly reduce inference costs; 2. Edge deployment feasibility—2B-4B models are suitable for edge devices, opening up possibilities for mobile applications and IoT scenarios; 3. Rapid domain adaptation—the controllable data construction mechanism allows rapid training of dedicated models by combining domain data.

OmniThoughtVis has limitations: the current research is based on the Qwen3-VL model family, and the effectiveness of other architectures needs further verification; the performance ceiling of the teacher model limits the performance of the student model. Future directions include exploring iterative distillation (using student models as new teachers), multi-teacher distillation (integrating knowledge from multiple models), and developing more efficient distillation algorithms to reduce training costs.

OmniThoughtVis provides a feasible path for the practical deployment of multimodal reasoning models. Through systematic data curation and knowledge distillation, it is expected to lower the deployment threshold without sacrificing reasoning capabilities. With the development of such technologies, powerful multimodal AI capabilities will become more accessible and easy to use, bringing transformative impacts to various practical applications.