SuperAPI AI Directory
Partner with us中文AI newsFavorites

AI developer platforms

3,642 results · Page 82 of 102

MM-EUREKA

AI developer platforms

MM-EUREKA is an open-source multimodal reasoning research project by ModalMinds that explores the frontiers of multimodal reasoning through rule-based reinforcement learning (RL). The project has released two multimodal reasoning models, MM-Eureka-Qwen-7B and MM-Eureka-Qwen-32B, and has open-sourced the complete training pipeline, code, data, and more. On the MMK12 evaluation set, the 7B model scores 66.1 and the 32B model scores 72.3; on MathVista, they reach 73.0 and 74.8 respectively, surpassing open-source models of the same scale and larger on some metrics. The project also proposes the CPGD reinforcement learning algorithm and the MM-PRM process reward model, and has publicly released the MMK12 dataset containing 15,000 multidisciplinary K12 samples and 2,000 multiple-choice questions for reasoning evaluation in subjects such as mathematics, physics, chemistry, and biology.

Read more

MMR1

AI developer platforms

MMR1 is an open-source research project for multimodal reasoning models (CVPR 2026), developed by LengSicong et al. The project addresses two major bottlenecks currently facing multimodal reasoning training—the lack of open, large-scale, high-quality long chain-of-thought (CoT) data, and the susceptibility of standard GRPO reinforcement learning to vanishing gradients under low reward variance, which leads to unstable fine-tuning—by proposing the Variance-Aware Sampling (VAS) method. Guided by the Variance Promotion Score (VPS), VAS combines outcome variance and trajectory diversity to select training prompts, thereby increasing reward variance, stabilizing policy optimization, and accelerating convergence. The project also open-sources the MMR1-SFT dataset of approximately 1.6 million long CoT cold-start trajectories, the MMR1-RL dataset of approximately 15,000 RL QA pairs, and multi-scale model checkpoints at 3B, 7B, and 32B, along with an end-to-end reproducible training pipeline. Evaluations show that MMR1-7B-RL achieves 7B-scale SOTA on multiple mathematical multimodal reasoning benchmarks. The project is licensed under Apache 2.0 and is a research preview for non-commercial use only.

Read more