OpenMMReasoner is a fully transparent two-stage multimodal reasoning solution that covers supervised fine-tuning (SFT) and reinforcement learning (RL). By carefully constructing high-quality datasets, this solution surpasses powerful baseline models in multiple multimodal reasoning benchmark tests, laying a solid empirical foundation for future large-scale multimodal reasoning research.

MultimodalTransformers

Emu3.5 is a native multimodal model developed by the Beijing Academy of Artificial Intelligence (BAAI). It can jointly predict the next state across vision and language, enabling coherent world modeling and generation. After end-to-end pre-training and large-scale reinforcement learning post-training, it demonstrates excellent performance in multimodal tasks.

MultimodalTransformers

Qwen3-VL is the most powerful vision-language model in the Qwen series to date. It has been comprehensively upgraded in text understanding and generation, visual perception and reasoning, context length, spatial and video dynamic understanding, and agent interaction capabilities. This model adopts a Mixture of Experts (MoE) architecture, providing excellent multimodal processing capabilities.

MultimodalGguf

This is a quantized version of the Qwen3-VL-2B-Instruct model. The llama.cpp tool and the imatrix method are used to generate model files with multiple quantization levels, which are convenient for efficient operation in different hardware environments. This model is a multimodal vision-language model with 2B parameters, supporting the interaction between images and text.

MultimodalGguf

Qwen3-VL is the most powerful vision-language model in the Tongyi series, with comprehensive upgrades in text understanding and generation, visual perception and reasoning, context length, spatial and video understanding capabilities, etc., and it has excellent multimodal interaction capabilities.

MultimodalTransformers

Apriel-1.5-15b-Thinker is a multimodal reasoning model in the ServiceNow Apriel SLM series, with 15 billion parameters. It can compete with models 10 times its size in text and image reasoning tasks. This model achieves excellent reasoning ability through a mid-term training scheme and can reach SOTA performance without image SFT training or reinforcement learning.

MultimodalTransformers

This is a GGUF format version converted from the Qwen3-VL-1B-Merged model, using the Q4_K_M quantization level. This model is a multimodal vision-language model with a scale of 1B parameters, supporting image understanding and text generation tasks.

MultimodalTransformersEnglish

InternVL3.5-4B is a medium-scale version in the open-source multimodal model series, achieving significant progress in generality, reasoning ability, and reasoning efficiency, and supporting new capabilities such as GUI interaction. The model adopts a cascaded reinforcement learning framework and visual resolution router technology to achieve efficient multimodal understanding and reasoning.

MultimodalTransformersOther

InternVL3_5-38B is a member of the open-source multimodal model InternVL3.5 series, which has made significant progress in versatility, reasoning ability, and reasoning efficiency. It supports multiple languages and can be applied to image-text-to-text tasks. It uses a cascaded reinforcement learning framework and visual resolution router technology to optimize performance.

MultimodalTransformersOther

InternVL3.5-4B is a medium-scale version of the open-source multimodal model series, containing 470 million parameters. It adopts an advanced cascaded reinforcement learning framework and visual resolution router technology, significantly improving the multimodal reasoning ability and efficiency.

MultimodalTransformersOther

InternVL3.5 is a new member of the open-source multimodal model family, significantly improving the versatility, reasoning ability, and reasoning efficiency of the InternVL series. It supports new functions such as GUI interaction and reaches the advanced level of open-source multimodal large language models.

MultimodalTransformersOther

Step-Audio 2 is an end-to-end multimodal large language model designed to meet the industry-level audio understanding and voice dialogue needs. It has advanced voice and audio understanding capabilities, intelligent voice dialogue functions, tool invocation, and multimodal retrieval enhanced generation capabilities, and has achieved leading performance in multiple audio understanding and dialogue benchmark tests.

MultimodalTransformers

Kuaishou Keye-VL is a cutting - edge multimodal large language model developed by the Kuaishou Keye team, which performs excellently in video understanding, visual perception, and reasoning tasks. The 1.5 version reaches a new height in video understanding, image perception, and reasoning ability through innovative fast - slow video encoding strategies, LongCoT cold - start data pipelines, and reinforcement learning training strategies, and supports an extended context length of up to 128k tokens.

MultimodalSafetensorsMultiple Languages

Gemma-SEA-LION-v4-27B-IT-GGUF is a quantized model based on Gemma-SEA-LION-v4-27B-IT, designed specifically for tasks in Southeast Asia. It has a large context length of 128K, multimodal understanding capabilities, and advanced function call features, and can run on ordinary laptops.

MultimodalTransformersMultiple Languages

Gemma 3 is a lightweight open-source multimodal model family launched by Google, built on the same research technology as Gemini. The 270M version is the smallest instruction-tuned model among them, supporting over 140 languages and having a 32K context length, suitable for text generation tasks in resource-constrained environments.

MultimodalTransformers

This is the GGUF quantized version of the XiaomiMiMo/MiMo-VL-7B-RL-2508 model, using the Q6_K quantization level. This model is a multimodal visual language model with a scale of 7B parameters, supporting joint understanding and generation tasks of images and text.

MultimodalTransformers

Intern-S1 is currently the most advanced open-source multimodal reasoning model. It combines powerful general task processing capabilities with excellent performance in a wide range of scientific tasks and can be comparable to leading closed-source commercial models. The model is continuously pre-trained on a 5T token dataset, of which more than 50% is professional scientific data. It has a dynamic word segmenter that can natively understand molecular formulas, protein sequences, and seismic signals.

MultimodalSafetensors

Intern-S1 is currently the most advanced open-source multimodal reasoning model, combining powerful general task capabilities with excellent performance on a wide range of scientific tasks, and can compete with leading closed-source commercial models.

MultimodalSafetensors

GLM-4.1V-9B-Base is an open-source vision-language foundation model developed by Zhipu AI. It has 9 billion parameters, focuses on multimodal reasoning capabilities, supports both Chinese and English, and can process images with up to 4K resolution and a context length of 64K.

MultimodalTransformersMultiple Languages

Patho-R1-7B is a multimodal reasoning model specifically designed for the field of pathology. It enhances the understanding ability of pathological diagnosis through a three-stage training pipeline (continuous pre-training, supervised fine-tuning, and reinforcement learning), and can effectively handle high-resolution pathological images and complex diagnostic reasoning tasks.

MultimodalTransformersEnglish