GPT-OSS-Safeguard-20B is a security reasoning model with 21 billion parameters built on GPT-OSS, specifically optimized for security-related text content classification and filtering tasks. This model supports custom security policies, provides a transparent reasoning process, and is suitable for security use cases such as input-output filtering of large language models and online content annotation.

Natural Language ProcessingTransformers

The INL architecture is a production-level neural architecture based on integral neuron dynamics, replacing traditional feed-forward network layers with iterative dynamics. It is suitable for various types of models such as large language models, vision transformers, and multi-modal models.

Natural Language ProcessingSafetensorsEnglish

LightOnOCR-1B-1025 is a compact end-to-end vision-language model specifically designed for optical character recognition and document understanding. It achieves state-of-the-art accuracy at the same weight level, while being faster and less costly than large general-purpose vision-language models.

MultimodalSafetensorsMultiple Languages

dParallel is a learnable parallel decoding method for large language models that achieves fast sampling by exploiting the intrinsic parallelism of the model. This method can significantly reduce the decoding steps while maintaining the model performance, achieving 8.5 - 10.5 times acceleration in benchmarks such as GSM8K and MBPP.

Natural Language ProcessingTransformers

Bee-8B is an advanced fully open-source multimodal large language model that focuses on data quality and aims to narrow the performance gap with proprietary models. By using the high-quality Honey-Data-15M corpus and the advanced data processing pipeline HoneyPipe, it demonstrates excellent performance in complex reasoning and other aspects.

MultimodalTransformersEnglish

This is the first - generation powerful vibe - code large language model, optimized to directly generate natural language and code completion based on loosely structured 'atmosphere coding' prompts. Compared with early models, it has lower prompt engineering overhead, smoother latent space interpolation, and is easier to guide the generation of usable code.

Natural Language ProcessingTransformersEnglish

Apertus is a fully open multilingual large language model developed by Swiss AI, offering two parameter scales of 7 billion and 8 billion. This model supports over 1000 languages, uses fully compliant and open training data, and its performance can rival that of closed-source models. Apertus was pre-trained on 15T tokens and adopts a phased curriculum training method, supporting a context length of up to 65,536 tokens.

Natural Language ProcessingTransformersOther

Apriel-1.5-15b-Thinker is a 15-billion parameter multimodal reasoning model developed by ServiceNow. It has the ability to reason with text and images, and its performance can rival models 10 times larger. It scored 52 points on the manual analysis index and performed excellently in enterprise domain benchmark tests.

MultimodalTransformers

Apertus is a fully open multilingual large language model developed by Swiss AI. It has parameter scales of 8 billion and 70 billion, supports over 1000 languages and long context processing, and only uses fully compliant open training data. Its performance can rival that of closed-source models.

Natural Language ProcessingTransformersOther

dParallel is a learnable parallel decoding method for large language models. It achieves fast sampling by exploring the intrinsic parallelism of the model, significantly reducing the decoding steps while maintaining performance, and achieving 8.5 - 10.5 times acceleration in benchmark tests such as GSM8K and MBPP.

Natural Language ProcessingTransformers

PatchFM is a basic model for univariate time series forecasting based on the Transformer architecture. It uses a block-by-block forecasting method, drawing on the training ideas of large language models to divide the time series into blocks for next-block forecasting. It supports multi-quantile output and uncertainty estimation.

Scientific ComputingSafetensors

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

This is a professional quantized version of the Qwen3-Omni model with 31.7B parameters. It uses advanced INT8+FP16 mixed-precision quantization technology, reducing memory usage by over 50%. It supports intelligent GPU/CPU hybrid inference, enabling large multimodal models to run efficiently on consumer-grade hardware.

MultimodalSafetensorsMultiple Languages

Apriel-1.5-15b-Thinker is a multimodal reasoning model developed by ServiceNow, with 15 billion parameters. It performs excellently in text and image reasoning tasks and can compete with models 10 times larger than it.

MultimodalTransformers

VLA-Adapter is a micro vision-language-action model trained on Libero-Spatial. It adopts the Prismatic-VLM architecture and only uses Qwen2.5-0.5B as the large language model backbone. This model outperforms open-source VLA models with larger parameter scales in robot benchmarks, achieving high-performance vision-language-action understanding and execution.

MultimodalSafetensorsEnglish

Isaac 0.1 is the first perceptual language model launched by PerceptronAI. It has 2 billion parameters and is designed for real-world applications. This model sets a new standard in efficiency, with performance reaching or exceeding that of models more than 50 times larger. It has powerful visual understanding and spatial reasoning capabilities.

MultimodalTransformers

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

Isaac-0.1 is the first open-source vision-language model launched by PerceptronAI. It has 2 billion parameters and is specifically designed for real-world applications. This model performs excellently in multimodal understanding and spatial reasoning, achieving performance that meets or exceeds models more than 50 times larger, setting a new efficiency standard.

MultimodalTransformers

LLaVA-OneVision-1.5 is a series of fully open-source large multimodal models that achieve advanced performance at a lower cost by training on native resolution images. This model demonstrates excellent performance in multiple multimodal benchmark tests, surpassing competitors such as Qwen2.5-VL.

MultimodalTransformers