DeepSeek-V3

DeepSeek-V3 is DeepSeek’s open mixture-of-experts large language model, whose official repository has 104,000+ stars on GitHub, and which reset expectations for what an openly released model could do.

Overview

DeepSeek-V3 is a 671-billion-parameter MoE model that activates only about 37 billion parameters per token, keeping inference costs far below its total size. It combines Multi-head Latent Attention with the DeepSeekMoE architecture, was pre-trained on 14.8 trillion tokens, and supports a 128K context window. On release it matched or beat leading closed models on many reasoning, coding and mathematics benchmarks while being trained at a fraction of their reported cost — a result that reshaped the open-model landscape and the market conversation around AI spending. This repository hosts the model code, inference demos and deployment guidance, including FP8 support and integrations with popular serving stacks. The repository saw updates in August 2025, and the model remains a reference point for efficient large-scale LLM design.

Key Features

• Efficient MoE design – 671B total parameters, roughly 37B active per token.
• MLA attention – Multi-head Latent Attention compresses the KV cache for cheap long-context inference.
• 128K context – handles long documents and extended reasoning traces.
• Frontier-class quality – strong results in reasoning, code generation, math and multilingual tasks.
• FP8 training and inference – engineered for modern GPU clusters and serving efficiency.
• Open release – weights and code available for research and commercial deployment.

💳 Pricing

The DeepSeek-V3 code in this repository is released under the MIT license and the model weights are openly downloadable; self-hosting costs only your infrastructure. DeepSeek also offers the model through its own API at published per-token rates.

Who Is It For

AI teams and researchers who want a frontier-class open model they can inspect, fine-tune and self-host, and anyone studying efficient MoE architecture design.