# Lightweight LLM

This is a quantized version based on the Pythia-70m model, specifically optimized for Wikipedia paragraph data, offering multiple quantization types to meet different resource requirements.

Qwen3-1.7B-4bit is a 4-bit quantized version of the Tongyi Qianwen 1.7B model, which has been converted to the MLX framework format for efficient operation on Apple Silicon devices.

Qwen3 is the latest generation of the Qwen series with 600 million parameters, supporting 32k context length and covering 119 languages

MiniCPM-S-1B-sft is a 1B-parameter language model optimized with activation sparsity techniques, achieving high-sparsity inference acceleration through the ProSparse method while maintaining performance comparable to the original model.

TxGemma is a series of lightweight, advanced open language models based on Gemma 2, specifically fine-tuned for therapeutic development. Available in 2B, 9B, and 27B sizes, it excels in processing information related to therapeutic modalities and targets.

TxGemma is a lightweight open-source language model based on Gemma 2, fine-tuned specifically for therapeutic development, available in 2B, 9B, and 27B sizes.

This is a statically quantized version of the knowledgator/Qwen-encoder-0.5B model, primarily designed for text encoding tasks.

A Qwen2.5-3B fine-tuned Arabic reasoning-specific language model, focused on enhancing Arabic language models' capabilities in logical reasoning and mathematical problem-solving.

This is a 4-bit quantized version of the QVikhr-2.5-1.5B-Instruct-SMPO model, optimized for the MLX framework, supporting Russian and English instruction understanding and generation tasks.

DeepSeek-R1-Distill-Llama-8B is a distilled large language model based on the Llama architecture, with a parameter scale of 8B, primarily designed for English text generation and comprehension tasks.

Phi-4 is an efficient small language model developed by Microsoft, focusing on high-performance inference under limited resources

A 1B-parameter quantized model based on Llama3.2 architecture, supporting text generation tasks with multiple quantization version options

A chat model with 4 billion parameters fine-tuned by H2O.ai, adjusted based on the Llama 2 architecture, supporting a context length of 8192

Cotype-Nano is a lightweight LLM designed to perform tasks with minimal resources. It is optimized for fast and efficient interaction with users, delivering high performance even under resource-constrained conditions.

A fine-tuned version of the NousResearch/Llama-3.2-1B model on the ultrachat_200k dataset, focusing on dialogue task optimization.

This is an ablated model based on mistralai/Mistral-Small-Instruct-2409, mainly used for text generation tasks.

Llama3.1-Neo-1B-100w is an efficient language model pruned to 1.4B parameters from Meta-Llama-3.1-8B-Instruct and fine-tuned using the LLM-Neo method (combining LoRA and knowledge distillation). The training data consists of 1 million samples from BAAI/Infinity-Instruct.

This is a version of the Llama-3-8b model quantized to INT4, using the QQQ quantization technique with a group size of 128 and optimized for hardware.

A 500M parameter dialogue fine-tuned model developed by H2O.ai, based on the Llama 2 architecture with Chinese dialogue support

Gemma is a lightweight open-source large language model series launched by Google, built on the same technology used to create Gemini models, suitable for various text generation tasks.

This is a small language model trained from scratch on the TinyChat dataset, aiming to achieve natural conversational responses with minimal model size.

An 8B parameter model based on the Llama 3 architecture, trained with various DPO datasets, focused on text generation tasks

LLaVA-Phi-3-mini is a fine-tuned LLaVA model based on Phi-3-mini-4k-instruct and CLIP-ViT-Large-patch14-336, specializing in image-to-text tasks.

Bllossom is a Korean-English bilingual language model based on Llama3, enhanced through comprehensive tuning to improve Korean language capabilities, expanding Korean vocabulary and optimizing Korean context processing.

WikiChat-v0.2 is a dialogue model currently under training, based on OpenOrca GPT-4 data, cosmopedia, and dolly15k datasets, supporting English text generation tasks.

A large language model built from scratch, with fully open-source implementations including tokenizer training, model initialization, pre-training, and instruction fine-tuning

This is a randomly initialized small model based on the DeepSeek-LLM-67B-Chat architecture, using float16 precision, primarily for text generation tasks.

Gemma-Ko is a Korean large language model developed based on Google's Gemma model, offering a 7B parameter version suitable for Korean and English text generation tasks.

TinyLLaVA is a small-scale large multimodal model framework that significantly reduces the number of parameters while maintaining high performance. The 3.1B version outperforms similar 7B-scale models in multiple benchmarks.

A sentiment analysis model fine-tuned on cryptocurrency news articles using the LoRA method based on the TinyLlama model

Llava-Phi2 is a multimodal implementation based on Phi2, combining vision and language processing capabilities, suitable for image-text-to-text tasks.

A 1.1B-parameter conversational language model fine-tuned on medical data, achieving an average 13.76% improvement on medical exam benchmarks

GGUF quantized version of MobileLLaMA-1.4B-Base, suitable for local deployment and inference

MobileLLaMA-1.4B-Base is a Transformer model with 1.4 billion parameters, designed for out-of-the-box deployment and trained on the RedPajama v1 dataset.

Cendol mT5-small Chat is a 300-million-parameter open-source generative large language model, fine-tuned for Indonesian, Sundanese, and Javanese instructions, suitable for single-turn dialogue scenarios.

TinyLlama is a small language model based on the LLaMA architecture with 1.1 billion parameters, fine-tuned using the alpaca-cleaned dataset.

The TinyLlama 1.5T checkpoint is a small language model based on 1.1B parameters, trained for answering questions.

Sheared-LLaMA-2.7B is a lightweight language model derived from Llama-2-7b through pruning and continued pretraining, consuming only a 50B token budget.

Sheared-LLaMA-1.3B is an efficient language model obtained through structured pruning and continual pre-training based on LLaMA-2-7B

Chinese-LLaMA-2-1.3B is a Chinese foundational model based on Meta's released Llama-2 model, expanded with a Chinese vocabulary and pre-trained in Chinese to enhance basic semantic understanding capabilities in Chinese.