Kimi - K2 - Thinking 3.825bit MLX is a quantized model for text generation. It achieves different perplexity performances in tests through different quantization methods. Among them, q3.825bit quantization can reach a perplexity of 1.256.

Natural Language ProcessingMlxEnglish

This project is a quantized version based on the MiniMaxAI/MiniMax-M2 model, aiming to make knowledge accessible to the public. It provides multiple model versions with different quantization levels and shows the perplexity performance indicators of each version.

Natural Language ProcessingGguf

MiniMax-M2 6.5bit MLX is a quantized version based on the MiniMax-M2 model, which performs excellently in text generation tasks and supports multiple quantization levels. Among them, the q6.5bit quantization can achieve a perplexity of 1.128 in testing, which is comparable to q8.

Natural Language ProcessingMlx

This is a collection of quantized versions of the inclusionAI/Ling-1T model, developed based on the ik_llama.cpp branch. It offers multiple quantization schemes to meet different memory and performance requirements. These quantized models provide excellent perplexity performance under a given memory footprint and are specifically optimized for the inference efficiency of the MoE architecture.

Natural Language ProcessingGguf

Qwen3-Coder-480B-A35B-Instruct is a text generation model based on the MLX library, specifically optimized for code generation tasks. This model offers multiple quantization levels. Under 6.5-bit quantization, it can achieve a perplexity performance comparable to that of 8-bit quantization, significantly reducing resource requirements while maintaining high-quality output.

Natural Language ProcessingMlx

A quantized version in GGUF format based on the DeepSeek-V3.1 model, using the ik_llama.cpp branch for optimal quantization. This model achieves the best perplexity performance under a specific memory footprint and offers various quantization schemes from IQ5_K to IQ1_S to meet different memory and performance requirements.

Natural Language ProcessingGgufOther

This is a GGUF format version optimized for Tigrinya based on Google's Gemma-3-4b model. It is specifically designed for Tigrinya text generation tasks and has achieved a perplexity of 2.5 on the Tigrinya news corpus.

Natural Language ProcessingGgufOther

Kimi-K2 Dynamic MLX is a text generation project built on the moonshotai/Kimi-K2-Instruct model, using the optimized MLX library to achieve efficient quantization performance. This model runs on a single M3 Ultra machine with 512GB RAM, supports multiple quantization methods, and shows excellent perplexity metrics in tests.

Natural Language ProcessingMlx

This is the imatrix quantized version of the Qwen3-235B-A22B-Thinking-2507 model. It uses the optimal quantization scheme from the ik_llama.cpp branch and achieves first-class perplexity performance under the given memory footprint. This model is a large language model with 235 billion parameters and supports multiple languages such as Chinese and English.

Natural Language ProcessingGguf

An 8B-parameter model based on Llama 3.1, supporting 1 million token context length, optimized for creative writing and role-playing with high stability and low perplexity characteristics.

Natural Language ProcessingGgufEnglish

A powerful text generation model based on Llama3, suitable for activities such as writing and role-playing, with high stability and low perplexity.

Natural Language ProcessingGgufEnglish

L3-Dark-Planet-8B-NEO-Imatrix-GGUF is an 8B parameter large language model based on the Llama3 architecture, specifically designed for creative writing, role-playing, and fictional creation. This model features low perplexity and high stability, and can run stably under various parameter settings, making it particularly suitable for scenarios requiring high-quality text generation.

Natural Language ProcessingGgufEnglish

A powerful text generation model designed for scenarios such as creative writing and role-playing, with excellent long text generation ability and low perplexity.

Natural Language ProcessingGgufEnglish

This is a language model with 1.3 billion parameters specifically designed for Azerbaijani. It is fine-tuned based on the mGPT-XL architecture and can efficiently process Azerbaijani text, achieving a perplexity of 5.37 on the validation set.

Natural Language ProcessingTransformersMultiple Languages

A Japanese language model with 3.6 billion parameters developed by LINE Corporation, based on the GPTNeoX architecture, supporting Japanese text generation tasks. The training data is approximately 650GB, and the perplexity on the Japanese C4 validation set is 7.50.

Natural Language ProcessingTransformersJapanese

A Japanese language model with 1.7 billion parameters developed by LINE Corporation, based on the GPT - 2 architecture and specifically trained for Japanese text generation tasks. The model was trained on approximately 650GB of Japanese corpora and achieved a perplexity of 8.57 on the Japanese C4 validation set.

Natural Language ProcessingTransformersJapanese

A 13-billion-parameter language model for Russian, pretrained on 300GB of multi-domain data with Russian perplexity around 8.8

Natural Language ProcessingTransformersMultiple Languages

A RoBERTa base model pre-trained on 1B tokens with a validation perplexity of 3.93, suitable for English text processing tasks.

Natural Language ProcessingJax

This is a GPT2 large model (762 million parameters) trained from scratch, focusing on Dutch language, with a perplexity of 15.1 on the clean Dutch mC4 dataset.

Natural Language ProcessingJaxOther

A text processing model fine-tuned based on bert-base-uncased, trained on an unknown dataset, with a validation loss of 2.4969 and perplexity of 12.14

Natural Language ProcessingTransformers