K-EXAONE 2.0 750B-A37B

K-EXAONE 2.0 750B-A37B is LG AI Research's mixture-of-experts language model with roughly 750 billion total parameters, 37 billion active parameters, and Apache 2.0 weights.

Platform: Replicate

Language ModelMixture of ExpertsOpen WeightsApache 2.0

Self-hosted accelerator cluster or FriendliAI-hosted API

Open weights under Apache 2.0; self-hosting costs vary and FriendliAI pricing is provider-specific

Commercial

🚀Function Overview

A very large self-hosted MoE model for multilingual reasoning, text generation, and tool-call experiments where operators can supply cluster-scale infrastructure.

Key Features

Official Hugging Face repository LGAI-EXAONE/K-EXAONE-2.0-750B-A37B
Approximately 750B total parameters and 37B activated parameters
256 experts with 8 experts selected per token in the published configuration
Ten languages listed in the official model metadata
Apache 2.0 model license
BF16 weights plus official FP8, NVFP4, and DSpark variants
Official LG AI Research web demo and a provider-specific FriendliAI API route

Use Cases

•Evaluating large multilingual MoE models on controlled internal workloads
•Trying the model through LG AI Research's web demo before planning cluster deployment
•Testing Korean and English reasoning or instruction-following tasks
•Building self-hosted text and tool-use experiments on cluster infrastructure
•Comparing full-precision and official quantized deployment variants

⚙️Input Parameters

messages

array

Chat messages formatted with the repository's published tokenizer chat template.

tools

array

Optional tool declarations supported by the published chat template; runtime behavior depends on the serving stack.

💡Usage Examples

Example 1

Input Parameters

{
  "messages": [
    {
      "role": "user",
      "content": "Summarize the deployment risks in this architecture proposal."
    }
  ]
}

Output Results

Generated text from the self-hosted K-EXAONE 2.0 serving stack.

Quick Actions

Use NowView Documentation

Technical Specifications

Hardware Type: Self-hosted accelerator cluster or FriendliAI-hosted API
Commercial Use: Supported
Pricing: Open weights under Apache 2.0; self-hosting costs vary and FriendliAI pricing is provider-specific
Platform: Replicate