Identifies Seeing but Not Thinking phenomenon in multimodal MoE models: vision tokens routed to reasoning experts cause distraction. Soft routing intervention improves Qwen3-VL-30B-A3B average from 59

Second only to Gemini 3 Pro (87.2) on multimodal understanding.

Qwen3.5 flagship MoE model (397B total, 17B active). Unified vision-language foundation model.

College-level visual reasoning. Led prior OpenAI models at launch.

Beats Qwen 3.5 27B and Gemma 4 31B on MMMU. Best 27B on multimodal reasoning.

8B open-weight multimodal model trained with GRPO+GDPO. Competitive with Gemini 2.5 Pro on DocVQA and chart understanding. New SOTA for open-weight VLMs on MMMU.

Slight improvement over Qwen 3.5 35B-A3B (81.4). Beats Claude Sonnet 4.5 (79.6).

10B open-source model rivals models 10-20x larger on MMMU. Trained with 1.4k RL iterations and PaCoRe.

Identifies Seeing but Not Thinking phenomenon in multimodal MoE models: vision tokens routed to reasoning experts cause distraction. Soft routing intervention improves Qwen3-VL-30B-A3B average from 59

Open-weight VLM competitive with GPT-5 mini (79.0) on multimodal understanding.

Strong vision-language at 9B. Beats GPT-5-Nano (75.8) and Gemini-2.5-Flash-Lite (73.4).

Instruction-tuned multimodal score.

8B open-source model trained with G2RPO surpasses GPT-4o on MMMU and 5 other VLM benchmarks.

Introduces Gaussian GRPO (G2RPO), replacing standard linear scaling in GRPO with non-linear distributional matching. OpenVLThinkerV2-7B achieves new SOTA for open-source 7B models on MMMU (71.6%), Mat

8B open-weight multimodal model trained with GRPO+GDPO. Competitive with Gemini 2.5 Pro on DocVQA and chart understanding. New SOTA for open-weight VLMs on MMMU.

Compact 15B open-weight multimodal reasoning model from Microsoft. Achieves competitive VLM performance with much less compute via careful data curation and dynamic-resolution encoders.

8B open-weight multimodal model trained with GRPO+GDPO. Competitive with Gemini 2.5 Pro on DocVQA and chart understanding. New SOTA for open-weight VLMs on MMMU.

Proposes RLSD combining self-distillation magnitude with RLVR direction. On Qwen3-VL-8B, achieves best avg accuracy across 5 multimodal reasoning benchmarks, outperforming GRPO by 2.32% on average.

4B end-to-end OCR model that ranks #1 on OmniDocBench among end-to-end models. Introduces "Layout-as-Thought" for structured layout representations. Outperforms Qwen3-VL-4B on ChartQA (+4.8) and Chart

Compact 15B open-weight multimodal reasoning model from Microsoft. Achieves competitive VLM performance with much less compute via careful data curation and dynamic-resolution encoders.

8B open-weight multimodal model trained with GRPO+GDPO. Competitive with Gemini 2.5 Pro on DocVQA and chart understanding. New SOTA for open-weight VLMs on MMMU.

4B end-to-end OCR model that ranks #1 on OmniDocBench among end-to-end models. Introduces "Layout-as-Thought" for structured layout representations. Outperforms Qwen3-VL-4B on ChartQA (+4.8) and Chart

Identifies Seeing but Not Thinking phenomenon in multimodal MoE models: vision tokens routed to reasoning experts cause distraction. Soft routing intervention improves Qwen3-VL-30B-A3B average from 59

Compact 15B open-weight multimodal reasoning model from Microsoft. Achieves competitive VLM performance with much less compute via careful data curation and dynamic-resolution encoders.

Compact 15B open-weight multimodal reasoning model from Microsoft. Achieves competitive VLM performance with much less compute via careful data curation and dynamic-resolution encoders.

Compact 15B open-weight multimodal reasoning model from Microsoft. Achieves competitive VLM performance with much less compute via careful data curation and dynamic-resolution encoders.

0.8B parameter VLM using Mamba-based SSM with cross-modality modulation instead of cross-attention. Outperforms MobileVLM V2 (1.7B) and MoE-LLaVA (2.2B) on several benchmarks despite being much smalle