GitHub – Deepseek-ai/DeepSeek-V3

We provide DeepSeek-V3, a strong Mixture-of-Experts (MoE) language design with 671B overall parameters with 37B activated for each token. To accomplish efficient inference and cost-efficient training, DeepSeek-V3 adopts Multi-head Latent Attention (MLA) and DeepSeekMoE architectures, which were completely confirmed in DeepSeek-V2. Furthermore, DeepSeek-V3 pioneers an auxiliary-loss-free method for load balancing and sets a multi-token prediction training objective for stronger efficiency. We pre-train DeepSeek-V3 on 14.8 trillion varied and top quality tokens, followed by Supervised Fine-Tuning and Reinforcement Learning stages to totally harness its abilities. Comprehensive evaluations reveal that DeepSeek-V3 surpasses other open-source designs and attains efficiency similar to leading closed-source designs. Despite its excellent performance, DeepSeek-V3 needs just 2.788 M H800 GPU hours for its full training. In addition, its training procedure is remarkably stable. Throughout the whole training procedure, we did not experience any irrecoverable loss spikes or perform any rollbacks.

2. Model Summary

Architecture: Innovative Load Balancing Strategy and Training Objective

– On top of the effective architecture of DeepSeek-V2, we leader an auxiliary-loss-free technique for load balancing, which decreases the efficiency destruction that emerges from encouraging load balancing.
– We examine a Multi-Token Prediction (MTP) objective and prove it advantageous to design efficiency. It can likewise be used for speculative decoding for inference acceleration.

Pre-Training: Towards Ultimate Training Efficiency

– We create an FP8 mixed accuracy training structure and, for the very first time, validate the expediency and efficiency of FP8 training on an exceptionally large-scale model.
– Through co-design of algorithms, structures, and hardware, we get rid of the communication bottleneck in cross-node MoE training, nearly attaining full computation-communication overlap.
This considerably enhances our training efficiency and decreases the training expenses, allowing us to even more scale up the model size without additional overhead.
– At a cost-effective expense of only 2.664 M H800 GPU hours, we complete the pre-training of DeepSeek-V3 on 14.8 T tokens, producing the currently strongest open-source base design. The subsequent training stages after pre-training need only 0.1 M GPU hours.

Post-Training: Knowledge Distillation from DeepSeek-R1

– We introduce an innovative approach to boil down reasoning capabilities from the long-Chain-of-Thought (CoT) design, specifically from one of the DeepSeek R1 series designs, into basic LLMs, especially DeepSeek-V3. Our pipeline elegantly includes the confirmation and reflection patterns of R1 into DeepSeek-V3 and notably enhances its reasoning performance. Meanwhile, we also maintain a control over the output design and length of DeepSeek-V3.

3. Model Downloads

The overall size of DeepSeek-V3 designs on Hugging Face is 685B, which consists of 671B of the Main Model weights and 14B of the Multi-Token Prediction (MTP) Module weights. **

To make sure optimal efficiency and versatility, we have actually partnered with open-source communities and hardware vendors to offer multiple methods to run the design in your area. For step-by-step assistance, have a look at Section 6: How_to Run_Locally.

For developers wanting to dive much deeper, we recommend checking out README_WEIGHTS. md for information on the Main Model weights and the Multi-Token Prediction (MTP) Modules. Please note that MTP assistance is presently under active development within the community, and we invite your contributions and feedback.

4. Evaluation Results

Base Model

Standard Benchmarks

Best outcomes are displayed in bold. Scores with a space not surpassing 0.3 are thought about to be at the very same level. DeepSeek-V3 accomplishes the finest efficiency on many standards, specifically on mathematics and code tasks. For more assessment details, please inspect our paper.

Context Window

Evaluation results on the Needle In A Haystack (NIAH) tests. DeepSeek-V3 carries out well throughout all context window lengths as much as 128K.

Chat Model

Standard Benchmarks (Models larger than 67B)

All designs are examined in a configuration that restricts the output length to 8K. Benchmarks including fewer than 1000 samples are times using differing temperature level settings to obtain robust last results. DeepSeek-V3 stands as the best-performing open-source model, and likewise displays competitive performance versus frontier closed-source designs.

Open Ended Generation Evaluation

English open-ended discussion evaluations. For AlpacaEval 2.0, we utilize the length-controlled win rate as the metric.

5. Chat Website & API Platform

You can chat with DeepSeek-V3 on DeepSeek’s official website: chat.deepseek.com

We also provide OpenAI-Compatible API at DeepSeek Platform: platform.deepseek.com

6. How to Run Locally

DeepSeek-V3 can be deployed locally utilizing the following hardware and open-source neighborhood software:

DeepSeek-Infer Demo: We supply an easy and lightweight demo for FP8 and BF16 inference.
SGLang: Fully support the DeepSeek-V3 model in both BF16 and FP8 reasoning modes, with Multi-Token Prediction coming quickly.
LMDeploy: Enables efficient FP8 and BF16 inference for local and cloud deployment.
TensorRT-LLM: Currently supports BF16 inference and INT4/8 quantization, with FP8 support coming soon.
vLLM: Support DeepSeek-V3 design with FP8 and BF16 modes for tensor parallelism and pipeline parallelism.
AMD GPU: Enables running the DeepSeek-V3 design on AMD GPUs by means of SGLang in both BF16 and FP8 modes.
Huawei Ascend NPU: Supports running DeepSeek-V3 on Huawei Ascend devices.
Since FP8 training is natively embraced in our structure, we just supply FP8 weights. If you require BF16 weights for experimentation, you can use the provided conversion script to perform the improvement.

Here is an example of converting FP8 weights to BF16:

Hugging Face’s Transformers has not been directly supported yet. **

6.1 Inference with DeepSeek-Infer Demo (example just)

System Requirements

Note

Linux with Python 3.10 only. Mac and Windows are not supported.

Dependencies:

Model Weights & Demo Code Preparation

First, clone our DeepSeek-V3 GitHub repository:

Navigate to the reasoning folder and set up dependences listed in requirements.txt. Easiest way is to use a bundle manager like conda or uv to produce a new virtual environment and set up the dependencies.

Download the model weights from Hugging Face, and put them into/ path/to/DeepSeek-V 3 folder.

Model Weights Conversion

Convert Hugging Face model weights to a specific format:

Run

Then you can chat with DeepSeek-V3:

Or batch reasoning on a given file:

6.2 Inference with SGLang (suggested)

SGLang currently supports MLA optimizations, DP Attention, FP8 (W8A8), FP8 KV Cache, and Torch Compile, delivering state-of-the-art latency and throughput performance amongst open-source frameworks.

Notably, SGLang v0.4.1 completely supports running DeepSeek-V3 on both NVIDIA and AMD GPUs, making it a highly versatile and robust service.

SGLang likewise supports multi-node tensor parallelism, enabling you to run this model on several network-connected makers.

Multi-Token Prediction (MTP) is in development, and development can be tracked in the optimization plan.

Here are the launch guidelines from the SGLang group: https://github.com/sgl-project/sglang/tree/main/benchmark/deepseek_v3

6.3 Inference with LMDeploy (advised)

LMDeploy, a flexible and high-performance reasoning and serving framework customized for big language models, now supports DeepSeek-V3. It offers both offline pipeline processing and online release capabilities, perfectly integrating with PyTorch-based workflows.

For extensive step-by-step directions on running DeepSeek-V3 with LMDeploy, please refer to here: InternLM/lmdeploy # 2960

6.4 Inference with TRT-LLM (advised)

TensorRT-LLM now supports the DeepSeek-V3 model, using accuracy options such as BF16 and INT4/INT8 weight-only. Support for FP8 is currently in progress and will be released soon. You can access the custom branch of TRTLLM particularly for DeepSeek-V3 support through the following link to experience the new functions directly: https://github.com/NVIDIA/TensorRT-LLM/tree/deepseek/examples/deepseek_v3.

6.5 Inference with vLLM (recommended)

vLLM v0.6.6 supports DeepSeek-V3 reasoning for FP8 and BF16 modes on both NVIDIA and AMD GPUs. Aside from basic strategies, vLLM offers pipeline parallelism enabling you to run this design on several devices connected by networks. For comprehensive assistance, please describe the vLLM directions. Please do not hesitate to follow the enhancement plan also.

6.6 Recommended Inference Functionality with AMD GPUs

In collaboration with the AMD group, we have attained Day-One assistance for AMD GPUs using SGLang, with complete compatibility for both FP8 and BF16 precision. For in-depth assistance, please describe the SGLang guidelines.

6.7 Recommended Inference Functionality with Huawei Ascend NPUs

The MindIE structure from the Huawei Ascend community has actually successfully adjusted the BF16 version of DeepSeek-V3. For step-by-step guidance on Ascend NPUs, please follow the guidelines here.

7. License

This code repository is accredited under the MIT License. Making use of DeepSeek-V3 Base/Chat models goes through the Model License. DeepSeek-V3 series (consisting of Base and Chat) supports industrial usage.