G²RPO is a novel reinforcement learning framework specifically designed for preference alignment in flow models, significantly improving the generation quality through a granular reward evaluation mechanism.

MultimodalDiffusers

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

QuestA is an innovative framework that enhances the reasoning ability of large language models through a question-enhanced approach. It integrates partial solutions during the reinforcement learning training process, significantly improving the model's performance on complex tasks such as mathematical reasoning, especially achieving optimal results on small-parameter models.

Natural Language ProcessingTransformers

MiMo Audio is an audio language model developed by Xiaomi, which demonstrates strong few-shot learning ability through large-scale pre-training. This model breaks through the limitations of traditional models that rely on fine-tuning for specific tasks and performs excellently in tasks such as speech intelligence and audio understanding, reaching an advanced level among open-source models.

Audio ProcessingSafetensors

MiMo Audio is an audio language model based on large-scale pre-training, achieving SOTA performance among open-source models in speech intelligence and audio understanding benchmark tests. This model demonstrates strong few-shot learning ability and can generalize to tasks not included in the training data, supporting various audio tasks such as speech conversion, style transfer, and speech editing.

Audio ProcessingSafetensors

DeepSeek-GRM-16B is a generative reward model based on Self-Principled Critique Tuning (SPCT), which can generate a transparent 'Principle → Critique → Score' evaluation process for query-responses and can be used in tasks such as reinforcement learning, evaluation, and data collection of large language models.

Natural Language ProcessingTransformersMultiple Languages

DINOv3 is a series of general visual foundation models developed by Meta AI. Without fine-tuning, it can outperform specialized state-of-the-art models in a wide range of visual tasks. This model uses self-supervised learning to generate high-quality dense features and performs excellently in various tasks such as image classification, segmentation, and depth estimation.

Computer VisionTransformersEnglish

DINOv3 is a series of general-purpose visual foundation models that can outperform specialized state-of-the-art models on a wide range of visual tasks without fine-tuning. The model uses self-supervised learning to generate high-quality dense features and performs excellently in various visual tasks, significantly surpassing previous self-supervised and weakly supervised foundation models.

Computer VisionTransformersEnglish

DINOv3 is a series of general visual foundation models that can outperform specialized state-of-the-art techniques in a wide range of visual tasks without fine-tuning. The model generates high-quality dense features through self-supervised learning and performs excellently in various visual tasks, significantly surpassing previous self-supervised and weakly supervised foundation models.

Computer VisionTransformersEnglish

GLM-4.1V-9B-Thinking is an open-source model developed based on the Vision-Language Model (VLM), aiming to explore the upper limit of the reasoning ability of vision-language models. By introducing the 'thinking paradigm' and reinforcement learning, it significantly improves the model's ability and reaches the leading level among VLMs with 10B parameters.

MultimodalTransformersMultiple Languages

GLM-4.1V-9B-Thinking is an open-source vision-language model based on the GLM-4-9B-0414 base model, specifically designed to explore the upper limit of the reasoning ability of vision-language models. It introduces the 'thinking paradigm' and uses reinforcement learning to significantly enhance its capabilities, performing excellently among VLMs with 10B parameters and rivaling or even surpassing models with larger parameters in multiple benchmark tasks.

MultimodalTransformersMultiple Languages

The Deep Ignorance Model Suite is a collection of 18 large language models with 6.9 billion parameters each. It aims to study methods of preventing models from learning unsafe technical capabilities (such as CBRN-related capabilities) by filtering pretrained data. This suite demonstrates that filtering data can effectively avoid the learning of bad knowledge while maintaining general performance and having anti-tampering capabilities.

Natural Language ProcessingTransformersEnglish

AReaL is an asynchronous reinforcement learning training system developed by Ant Group, designed specifically for large inference models, supporting fast training and cutting-edge performance.

Natural Language ProcessingTransformers

Llama-3.1-8B-Instruct-RM-RB2 is one of the reward models released by the Allen Institute for Artificial Intelligence. It is fine-tuned based on Llama-3.1-8B-Instruct and is used to evaluate and optimize the preference learning of generative models.

Natural Language ProcessingTransformersEnglish

VeriReason is a Verilog RTL code generation method that combines reinforcement learning with testbench feedback, significantly improving the performance of pre-trained models in the field of hardware design.

Natural Language ProcessingTransformers

VeriReason is a reinforcement learning method that utilizes testbench feedback to improve the performance of pre-trained models in Verilog RTL code generation.

Natural Language ProcessingTransformers

R1-Code-Interpreter is the first framework that uses multi-round supervised fine-tuning and reinforcement learning to train large language models for step-by-step code reasoning. It supports 144 different reasoning and planning tasks, enabling the model to autonomously decide when and how to invoke code.

Natural Language ProcessingTransformers

UniME is a general embedding learning framework based on multimodal large models, significantly enhancing multimodal embedding capabilities through text discriminative knowledge distillation and hard negative sample-enhanced instruction tuning strategies.

MultimodalTransformersEnglish

SkyReels V2 is the world's first unlimited-length cinematic generation model employing a diffusion-forcing framework, integrating multimodal large language models, multi-stage pretraining, reinforcement learning, and diffusion-forcing technology for comprehensive optimization.

MultimodalDiffusers

OctoThinker is a reinforcement learning - friendly base language model built on the Llama - 3 family of models, focusing on the research of reinforcement learning extension with mid - term training incentives. This model with a 3B parameter scale has important application value in the field of natural language processing and is particularly suitable for text generation tasks.

Natural Language ProcessingSafetensorsEnglish