---
title: "DeepSeek V2 — LLM profile"
provider: "DeepSeek"
version: "v2"
model_type: "llm"
context_window: "131072"
parameters: "236B"
url: "https://lforla.org/models/deepseek-v2"
markdown_twin: "https://lforla.org/markdown/models/deepseek-v2.md"
---
# DeepSeek V2

## Overview

DeepSeek-V2 is a strong, economical, and efficient Mixture-of-Experts (MoE) language model released by DeepSeek-AI in May 2024, designed to lower training and inference costs while maintaining high performance.

## Capabilities

- 128k-token context window
- Strong performance across general and code benchmarks
- Efficient inference with a fine-grained MoE architecture
- Open weights

## Technical details

- 236 billion total parameters with 21 billion activated per token (MoE)
- Introduces Multi-head Latent Attention (MLA) and DeepSeekMoE for economical training
- 67% cost reduction compared to DeepSeek-V1 during training

## References

- DeepSeek-V2: A Strong, Economical, and Efficient Mixture-of-Experts Language Model — [arXiv:2405.04434](https://arxiv.org/abs/2405.04434)