MiniCPM4 is an efficient large language model designed specifically for edge devices. Through systematic innovation, it achieves extreme efficiency improvements in four dimensions: model architecture, training data, training algorithm, and inference system. It can achieve over 5 times faster generation speed on edge chips.

Natural Language ProcessingTransformersMultiple Languages

An efficient inference model fine-tuned based on Qwen 30B-A3B (MOE), achieving nearly double the speed improvement by reducing the number of experts, supporting multiple quantization formats and a 40K context length.

Natural Language ProcessingTransformers

This is a Mistral-7B instruction fine-tuned model optimized using Unsloth and Huggingface TRL library, achieving 2x training speed improvement

Natural Language ProcessingTransformersEnglish

The Phi-4-mini-instruct model undergoes float8 dynamic activation and weight quantization via torchao, achieving 36% VRAM reduction and 15-20% speed improvement on H100 with minimal accuracy impact.

Natural Language ProcessingTransformersOther

Fine-tuned version based on Qwen2.5-VL-3B model, trained using Unsloth and Huggingface TRL library, achieving 2x inference speed improvement

MultimodalTransformersEnglish

Currently the fastest Danish ASR model, a distilled version of hviske-v2, achieving approximately 4x speed improvement while maintaining the same accuracy.

Optimized Qwen2 model based on Unsloth and Huggingface TRL library, achieving 2x inference speed improvement

Natural Language ProcessingTransformersEnglish

Llama model trained with Unsloth and Huggingface TRL library, achieving 2x inference speed improvement

Natural Language ProcessingTransformersEnglish

A fine-tuned version of the Qwen2-VL-7B model, trained using Unsloth and Huggingface's TRL library, achieving a 2x speed improvement.

MultimodalTransformersEnglish

A fine-tuned version of the Qwen2-VL-7B model, trained using Unsloth and Huggingface TRL library, achieving 2x inference speed improvement.

MultimodalTransformersEnglish

INT4 quantized version of FLUX.1-Fill-dev, capable of filling regions in existing images based on text descriptions, providing approximately 4x memory savings and 2-3x speed improvements.

Computer VisionDiffusersEnglish

A multimodal model fine-tuned based on Qwen2.5-VL-7B-Instruct, utilizing the Unsloth acceleration framework and TRL library for training, achieving a 2x speed improvement

MultimodalTransformersEnglish

An 8-layer embedding model distilled from BAAI/bge-m3, achieving 2.5x speed improvement while maintaining retrieval performance

Natural Language ProcessingSafetensors

FastHunyuan is the accelerated version of HunyuanVideo, requiring only 6 diffusion sampling steps to generate high-quality videos, achieving approximately an 8x speed improvement compared to the original version.

MultimodalSafetensors

An instruction-following model fine-tuned based on unsloth/Qwen2-7B-Instruct, trained using Unsloth and TRL libraries with 2x speed improvement

Natural Language ProcessingTransformersMultiple Languages

Kotoba-Whisper is a Japanese automatic speech recognition distilled model developed by Asahi Ushio in collaboration with Kotoba Technologies, based on Whisper large-v3 distillation, achieving a 6.3x inference speed improvement.

Audio ProcessingTransformersJapanese

A 4-bit quantized version based on Meta-Llama-3.1-8B-Instruct, trained using Unsloth and Huggingface TRL libraries, achieving 2x speed improvement.

Natural Language ProcessingGgufEnglish

A Python code generation model fine-tuned based on unsloth/Phi-3-mini-4k-instruct-bnb-4bit, trained using Unsloth and TRL libraries with 2x speed improvement

Natural Language ProcessingGgufEnglish

YOLOv10 is a real-time end-to-end object detection model proposed by Tsinghua University, with significant improvements in both speed and accuracy.

Computer VisionSafetensors

A text generation model fine-tuned based on unsloth/llama-3-8b-bnb-4bit, trained using Unsloth and TRL libraries with 2x speed improvement

Natural Language ProcessingTransformersEnglish