Track, rank and evaluate open LLMs and chatbots
Quantize a model for faster inference
Generate and view leaderboard for LLM evaluations
Browse and evaluate ML tasks in MLIP Arena
View LLM Performance Leaderboard
Evaluate model predictions with TruLens
Display and submit language model evaluations
Browse and evaluate language models
Push a ML model to Hugging Face Hub
Track, rank and evaluate open LLMs and chatbots
Browse and submit LLM evaluations
Compare model weights and visualize differences
Merge Lora adapters with a base model
The Low-bit Quantized Open LLM Leaderboard is a comprehensive tool designed to track, rank, and evaluate open large language models (LLMs) and chatbots, with a focus on low-bit quantization. It helps users explore and compare different LLMs, providing insights into their performance under quantized conditions. This leaderboard is particularly useful for developers, researchers, and enthusiasts looking to optimize model efficiency without compromising accuracy.
• Quantized Benchmarking: Evaluates models using low-bit quantization to reduce memory usage and increase inference speed.
• Model Comparison: Enables side-by-side comparison of different LLMs based on their quantized performance.
• Multi-bit Support: Covers models quantized to 4-bit, 8-bit, and other low-bit representations.
• Real-time Updates: Provides the latest rankings and performance metrics as new models emerge.
• Customizable Filters: Allows users to filter models by specific criteria like quantization bit, model size, or benchmark results.
• Performance Metrics: Displays key metrics such as accuracy, inference speed, and memory usage for each model.
What is low-bit quantization?
Low-bit quantization is a technique to reduce the precision of model weights, typically from 32-bit floating-point numbers to 4-bit or 8-bit integers, enabling faster inference and smaller model sizes.
Which quantization bits are supported?
The leaderboard supports models quantized to 4-bit, 8-bit, and other low-bit representations, ensuring a wide range of optimized models are available for comparison.
How are models ranked?
Models are ranked based on their performance in quantized benchmarks, considering metrics like accuracy, inference speed, and memory efficiency. Rankings are updated in real-time as new models are added.