RND1 is an experimental diffusion language model with 30 billion parameters, adopting a sparse mixture of experts architecture. This model is converted from a pre-trained autoregressive base model, supporting diffusion-based text generation. Only 3 billion parameters are activated for each token, achieving a balance between computational efficiency and model capacity.

Natural Language ProcessingTransformers

Ming-UniVision is a large multimodal language model that first integrates continuous visual representations into the next-token prediction framework, unifying vision and language under a single autoregressive paradigm without discrete quantization or modality-specific heads. This model supports joint image understanding and generation, converges faster in vision-language training, and also supports multi-round context visual tasks.

MultimodalTransformersMultiple Languages

SDAR is a new type of large language model that integrates autoregressive and discrete diffusion modeling strategies, combining the advantages of efficient training of AR models and parallel inference of diffusion models. It performs comparably to SOTA open - source AR models on general tasks and excels in scientific reasoning tasks, becoming the most powerful diffusion language model.

Natural Language ProcessingTransformers

The NVIDIA Qwen3-235B-A22B FP4 model is a quantized version of the Alibaba Cloud Qwen3-235B-A22B model. It adopts an optimized Transformer architecture and is an autoregressive language model. This model reduces the parameters from 16 bits to 4 bits through FP4 quantization technology, reducing the disk size and GPU memory requirements by approximately 3.3 times while maintaining high accuracy and performance.

Natural Language ProcessingSafetensors

SongBloom is an innovative full-length song generation framework that adopts an interleaved paradigm of autoregressive sketch drawing and diffusion-based refinement, combining the high fidelity of diffusion models and the scalability of language models to generate high-quality full-length songs.

Show-o2 is an improved native unified multimodal model that utilizes autoregressive modeling and flow matching techniques to support unified understanding and generation of text, image, and video modalities.

MultimodalDiffusers

A quantized version of the DeepSeek R1 0528 model from DeepSeek AI, an autoregressive language model based on an optimized Transformer architecture, which can be used for commercial and non-commercial purposes.

Natural Language ProcessingSafetensors

ReasonGen-R1 is an autoregressive image generation model that integrates chain-of-thought reasoning. It enhances the logic and quality of image generation through SFT and RL.

MultimodalTransformers

Dimple is the first discrete diffusion multimodal large language model (DMLLM) that combines autoregressive and diffusion training paradigms. After training on the same dataset as LLaVA-NEXT, it outperforms LLaVA-NEXT-7B by 3.9%.

MultimodalTransformersEnglish

PARD is a high-performance speculative decoding method that can convert autoregressive draft models into parallel draft models at low cost, significantly accelerating the inference of large language models.

Natural Language ProcessingTransformers

Autoregressive-32B is a Multiverse-32B baseline model built based on autoregressive modeling, providing strong support for text generation tasks.

Natural Language ProcessingTransformers

AuriStream-1B is a biologically inspired GPT-style autoregressive Transformer model specifically designed to predict cochlear markers in long speech contexts. The model uses a long context window of approximately 20 seconds (4096 tokens) and was trained on the LibriLight dataset (approximately 60,000 hours) for about 500,000 steps. It can learn rich time-aligned representations and generate speech continuations.

Audio ProcessingTransformersEnglish

XGLM-564M is a multilingual autoregressive language model with 564 million parameters, trained on a balanced corpus of 30 languages totaling 500 billion subwords.

Natural Language ProcessingJaxMultiple Languages

SkyReels V2 is the first open-source video generation model to adopt an autoregressive diffusion-forced architecture, supporting unlimited-length movie generation and achieving state-of-the-art performance among public models.

Computer VisionDiffusers

SkyReels V2 is an unlimited-length cinematic generation model that employs an autoregressive diffusion-forced architecture, supporting high-quality video generation.

Computer VisionSafetensors

SkyReels V2 is an unlimited-length cinematic generation model that employs an autoregressive diffusion-forced architecture, supporting high-resolution video generation.

MultimodalDiffusers

SkyReels V2 is an unlimited-length cinematic generation model that adopts an autoregressive diffusion-forced architecture, supporting text-to-video and image-to-video tasks, capable of generating high-quality long video content.

MultimodalDiffusers

SkyReels V2 is an open-source video generation model that adopts an autoregressive diffusion forced architecture, supporting unlimited-length movie generation and achieving state-of-the-art performance among public models.

MultimodalSafetensors

SkyReels V2 is an unlimited-length cinematic generation model that adopts an autoregressive diffusion force architecture, supports 720P HD video generation, and achieves SOTA performance among public models.

Computer VisionSafetensors

SkyReels V2 is the first open-source video generation model adopting an autoregressive diffusion forced architecture, supporting unlimited-length movie generation, achieving state-of-the-art performance among public models.

Computer VisionSafetensors