Description
DeepSeek-V3 is a large language model released by DeepSeek AI as a Mixture-of-Experts (MoE) architecture: 671 billion parameters in total, with only 37 billion activated for any given token, which keeps inference cheaper than running the full parameter count on every request. It builds on the architecture introduced in DeepSeek-V2, carrying forward Multi-head Latent Attention and the DeepSeekMoE design, and adds an auxiliary-loss-free strategy for balancing load across experts, which the team says reduces the performance cost that usually comes with load balancing.
Read More
Training used 14.8 trillion tokens of pre-training data, followed by supervised fine-tuning and reinforcement learning stages. DeepSeek reports the full pre-training run cost about 2.664 million H800 GPU hours, with later training stages adding only about 0.1 million more, for roughly 2.788 million GPU hours total, and describes the training process as stable, without irrecoverable loss spikes or rollbacks. Part of that efficiency comes from an FP8 mixed-precision training framework the team validated at this scale, alongside engineering work to reduce communication overhead across GPU nodes during Mixture-of-Experts training. A separate post-training step distills reasoning behavior from DeepSeek-R1, a long-chain-of-thought model in DeepSeek's lineup, into DeepSeek-V3, aiming to improve its reasoning while keeping its output style controlled.
DeepSeek-V3 supports a 128K context length and, on the benchmarks DeepSeek publishes, performs comparably to leading closed-source models while outperforming other open-source models it's compared against. The model and its base checkpoint, DeepSeek-V3-Base, are both downloadable from Hugging Face.
For anyone who doesn't want to self-host, DeepSeek runs a chat website where the model can be used directly. For self-hosting, DeepSeek-V3 is supported by several open-source inference engines, including SGLang, vLLM, LMDeploy, and TensorRT-LLM, with SGLang also enabling deployment on AMD GPUs. Running the demo inference script locally requires a Linux system with Python 3.10 (Mac and Windows aren't supported for that path) and, for the full model, enough combined GPU memory across multiple machines to hold 671 billion parameters. The code is released under the MIT license, while the model weights carry a separate model license agreement. This is a model for developers and researchers with the infrastructure to serve a large open-weight model, not a plug-and-play consumer app.
- CategoryAI Models & LLMs
- Pricing ModelOpen source
- DeploymentSelf-hosted
- PlatformsLinux
DeepSeek-V3's Core Features
671B-parameter Mixture-of-Experts model with 37B parameters activated per token
Multi-head Latent Attention and DeepSeekMoE architecture carried over and refined from DeepSeek-V2
Auxiliary-loss-free load balancing strategy across experts
Multi-token prediction training objective, also usable for speculative decoding at inference
FP8 mixed-precision training framework validated at large scale
128K context length
Post-training distillation of reasoning behavior from DeepSeek-R1 into the base model
Supported by major open-source inference engines: SGLang, vLLM, LMDeploy, and TensorRT-LLM, plus AMD GPUs via SGLang
Getting Started with DeepSeek-V3
Try it online: Chat with DeepSeek-V3 on DeepSeek's chat website without installing anything
Download the weights: Get DeepSeek-V3 or DeepSeek-V3-Base from Hugging Face
Pick an inference engine: Choose SGLang, vLLM, LMDeploy, or TensorRT-LLM based on your hardware
Convert weights if needed: Run the provided conversion script for the demo inference path
Run the model: Launch inference across enough GPUs and, for the full model, across multiple nodes
DeepSeek-V3's Use Cases
- Self-hosted LLM deployment
- MoE architecture research
- High-throughput inference serving
