Chonky is a multilingual paragraph segmentation model based on DistilBERT, fine-tuned on nearly 11 billion tokens from over 34 million Wikipedia articles. It can segment any natural language text into semantic chunks and supports 104 languages.

Natural Language ProcessingTransformersMultiple Languages

This model is a fine-tuned version based on meta-llama/Llama-3.3-70B-Instruct. It is trained using the mlabonne/FineTome-100k dataset, which contains 100k token data. The model is fine-tuned using the Unsloth and Huggingface TRL libraries and supports English language processing.

Natural Language ProcessingTransformersEnglish

The top model in the aquif-3.5 series, with advanced inference capabilities and a large context window of 1 million tokens, and has excellent performance in multiple benchmark tests. The comprehensive AAII score reaches 60 points.

Natural Language ProcessingTransformersMultiple Languages

The aquif-3.5 series is the masterpiece released on November 3, 2025, including two versions, Plus and Max, which provide advanced inference capabilities and an unprecedented 1 million token context window, achieving state-of-the-art performance in their respective categories.

Natural Language ProcessingTransformersMultiple Languages

WavJEPA is an audio foundation model based on the waveform joint embedding prediction architecture. It uses advanced semantic representation learning to address the deficiencies in speech unit or token-level representation learning. It significantly outperforms the state-of-the-art time-domain audio foundation models in many downstream benchmark tasks while requiring significantly less computational resources.

Audio ProcessingTransformers

Kimi Linear is an efficient hybrid linear attention architecture that outperforms traditional full attention methods in short context, long context, and reinforcement learning scenarios. It optimizes attention computation through the Kimi Delta Attention (KDA) mechanism, significantly improving performance and hardware efficiency, and is particularly good at handling long context tasks up to 1 million tokens.

Natural Language ProcessingTransformers

This is the third version of the monolingual encoder-decoder language model released by the HPLT project, specifically designed for German. It is a text-to-text transformer model trained based on the denoising objective, following the NorT5 architecture settings, and using a tokenizer trained with German HPLT data.

Natural Language ProcessingPytorchMultiple Languages

This is the GGUF quantized version of the AvitoTech/avibe model. It is based on the Qwen3-8B base model and undergoes Russian supervised fine-tuning. A custom tokenizer is used to optimize the Russian processing ability, and it is specifically designed for Russian text generation tasks.

Natural Language ProcessingGgufMultiple Languages

Svara-TTS is an open-source multilingual text-to-speech model for Indian languages, supporting 19 languages (18 Indian languages + Indian English). This model is built based on the Orpheus-style discrete audio token method, aiming to achieve clear, expressive, and low-latency speech synthesis on ordinary GPUs/CPUs.

Audio ProcessingTransformersMultiple Languages

This is a sequence-to-sequence model specifically designed to compress short user prompts (≤64 tokens) and can achieve low-latency operation of less than 100 milliseconds on modern GPUs. It serves as a lightweight preprocessing stage to support high-capacity large language models.

Natural Language ProcessingTransformers

Qwen3-Next-80B-A3B-Thinking-1M-qx64n-mlx is a large language model that adopts a mixture-of-experts architecture and Deckard mixed precision quantization technology. It has 80 billion parameters and a context length of 1M tokens. This model performs excellently in science and technology reasoning and long text processing, showing a 20 - 35% improvement in cognitive benchmark tests compared to instruction models.

Natural Language ProcessingMlx

The Llama Nemotron Embedding 1B model is an embedding model developed by NVIDIA, optimized for multilingual and cross - language text question - answering retrieval. It supports 26 languages, can handle documents up to 8192 tokens long, and can significantly reduce data storage requirements through dynamic embedding sizes.

Natural Language ProcessingTransformersOther

The Ming-flash-omni preview version is a multimodal large model built on the Ling-Flash-2.0 sparse mixture of experts (MoE) architecture, with a total of 100B parameters and only 6B parameters activated per token. This model is a comprehensive upgrade based on Ming-Omni, showing significant improvements in multimodal understanding and generation, especially in speech recognition, image generation, and segmentation editing.

MultimodalDiffusersEnglish

This is a compact Spanish-to-English translation model optimized for accuracy and efficiency. The model size is 386MB, supporting a context length of 4096 tokens, and can efficiently and accurately complete translation tasks with an average response time of approximately 245ms.

Natural Language ProcessingGgufMultiple Languages

LFM2-8B-A1B is an 8-bit quantized MLX build version optimized for Apple Silicon chips, adopting the Mixture of Experts (MoE) architecture. It has approximately 8 billion total parameters, with about 1 billion parameters activated per token, and supports fast inference on the device side.

Natural Language ProcessingMlxEnglish

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

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

PaDT Pro 3B is a multimodal large language model based on Qwen2.5VL-3B. It adopts a unified paradigm of decodable block tokens and can directly generate text and visual outputs, achieving advanced performance in multiple visual perception and understanding tasks.

MultimodalTransformersMultiple Languages

GLM 4.6 is a quantized model optimized for a configuration of 128GB memory + single GPU. It uses the IQ_K quantization method, providing better quality and performance than the mainstream llama.cpp at the same size. This model needs to be used with ik_llama.cpp. On a configuration of 128GB dual-channel DDR5 memory, a single CCD Ryzen 7000 processor + a single 3090 graphics card, the text generation speed can reach approximately 6.8 tokens per second.

Natural Language ProcessingGgufMultiple Languages

MingTok is an innovative visual tokenizer that achieves unified visual understanding and generation through a continuous latent space. It does not require vector quantization and performs excellently in image reconstruction and autoregressive convergence speed while preserving semantic and perceptual fidelity.

Computer VisionTransformersMultiple Languages