GitHub – Deepseek-ai/DeepSeek-V3

We present DeepSeek-V3, a strong Mixture-of-Experts (MoE) language design with 671B overall criteria with 37B triggered for each token. To accomplish effective inference and cost-effective training, DeepSeek-V3 embraces Multi-head Latent Attention (MLA) and DeepSeekMoE architectures, which were completely confirmed in DeepSeek-V2. Furthermore, DeepSeek-V3 leaders an auxiliary-loss-free technique for load balancing and sets a multi-token forecast training goal for stronger performance. We pre-train DeepSeek-V3 on 14.8 trillion diverse and high-quality tokens, followed by Supervised Fine-Tuning and Reinforcement Learning phases to fully harness its abilities. Comprehensive examinations reveal that DeepSeek-V3 surpasses other open-source designs and achieves efficiency similar to leading closed-source models. Despite its exceptional efficiency, DeepSeek-V3 requires just 2.788 M H800 GPU hours for its complete training. In addition, its training process is incredibly steady. Throughout the entire 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 efficient architecture of DeepSeek-V2, we pioneer an auxiliary-loss-free strategy for load balancing, which minimizes the performance destruction that develops from motivating load balancing.
– We examine a Multi-Token Prediction (MTP) objective and prove it useful to design performance. It can also be utilized for speculative decoding for reasoning velocity.

Pre-Training: Towards Ultimate Training Efficiency

– We create an FP8 combined accuracy training framework and, for the first time, confirm the expediency and efficiency of FP8 training on an exceptionally massive design.
– Through co-design of algorithms, structures, and hardware, we conquer the communication traffic jam in cross-node MoE training, nearly attaining full computation-communication overlap.
This significantly improves our training effectiveness and lowers the training expenses, enabling us to even more scale up the model size without extra overhead.
– At an economical expense of just 2.664 M H800 GPU hours, we finish the pre-training of DeepSeek-V3 on 14.8 T tokens, producing the presently 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 thinking abilities from the long-Chain-of-Thought (CoT) model, specifically from among the DeepSeek R1 series models, into basic LLMs, especially DeepSeek-V3. Our pipeline elegantly includes the confirmation and reflection patterns of R1 into DeepSeek-V3 and significantly enhances its reasoning performance. Meanwhile, we also preserve a control over the output style and length of DeepSeek-V3.

3. Model Downloads

The total size of DeepSeek-V3 models 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 ideal efficiency and flexibility, we have actually partnered with open-source communities and hardware suppliers to provide multiple methods to run the design in your area. For detailed guidance, examine out Section 6: How_to Run_Locally.

For designers aiming to dive 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 advancement within the neighborhood, and we invite your contributions and feedback.

4. Evaluation Results

Base Model

Standard Benchmarks

Best results are revealed in vibrant. Scores with a gap not surpassing 0.3 are thought about to be at the exact same level. DeepSeek-V3 attains the best performance on the majority of criteria, specifically on mathematics and code jobs. For more evaluation details, please examine our paper.

Context Window

Evaluation results on the Needle In A Haystack (NIAH) tests. DeepSeek-V3 performs well throughout all context window lengths up to 128K.

Chat Model

Standard Benchmarks (Models larger than 67B)

All designs are evaluated in a setup that restricts the output length to 8K. Benchmarks containing fewer than 1000 samples are tested several times using differing temperature settings to derive robust outcomes. DeepSeek-V3 stands as the best-performing open-source model, and also shows competitive performance against frontier closed-source designs.

Open Ended Generation Evaluation

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

5. Chat Website & API Platform

You can talk with DeepSeek-V3 on DeepSeek’s main site: chat.deepseek.com

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

6. How to Run Locally

DeepSeek-V3 can be deployed in your area using the following hardware and open-source community software application:

DeepSeek-Infer Demo: We offer an easy and light-weight demo for FP8 and BF16 inference.
SGLang: Fully support the DeepSeek-V3 design in both BF16 and FP8 reasoning modes, with Multi-Token Prediction coming quickly.
LMDeploy: Enables efficient FP8 and BF16 reasoning for regional and cloud implementation.
TensorRT-LLM: Currently supports BF16 reasoning and INT4/8 quantization, with FP8 assistance coming quickly.
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 through 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 framework, we only supply FP8 weights. If you require BF16 weights for experimentation, you can use the provided conversion script to perform the transformation.

Here is an example of converting FP8 weights to BF16:

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

6.1 Inference with DeepSeek-Infer Demo (example only)

System Requirements

Note

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

Dependencies:

Model Weights & Demo Code Preparation

First, clone our DeepSeek-V3 GitHub repository:

Navigate to the inference folder and set up reliances listed in requirements.txt. Easiest way is to use a plan supervisor like conda or uv to develop a brand-new virtual environment and install the dependencies.

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

Model Weights Conversion

Convert Hugging Face design weights to a particular format:

Run

Then you can talk with DeepSeek-V3:

Or batch inference on a given file:

6.2 Inference with SGLang (recommended)

SGLang currently supports MLA optimizations, DP Attention, FP8 (W8A8), FP8 KV Cache, and Torch Compile, delivering modern latency and throughput efficiency among open-source frameworks.

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

SGLang also supports multi-node tensor parallelism, enabling you to run this model on numerous network-connected devices.

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

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

6.3 Inference with LMDeploy (recommended)

LMDeploy, a flexible and high-performance reasoning and serving framework tailored for large language models, now supports DeepSeek-V3. It provides both offline pipeline processing and online release capabilities, seamlessly incorporating with PyTorch-based workflows.

For thorough detailed directions on running DeepSeek-V3 with LMDeploy, please describe here: InternLM/lmdeploy # 2960

6.4 Inference with TRT-LLM (advised)

TensorRT-LLM now supports the DeepSeek-V3 design, providing accuracy choices such as BF16 and INT4/INT8 weight-only. Support for FP8 is currently in development and will be launched quickly. You can access the custom branch of TRTLLM specifically for DeepSeek-V3 assistance through the following link to experience the brand-new features directly: https://github.com/NVIDIA/TensorRT-LLM/tree/deepseek/examples/deepseek_v3.

6.5 Inference with vLLM (advised)

vLLM v0.6.6 supports DeepSeek-V3 reasoning for FP8 and BF16 modes on both NVIDIA and AMD GPUs. Aside from basic techniques, vLLM uses pipeline parallelism permitting you to run this design on several machines connected by networks. For comprehensive guidance, please describe the vLLM directions. Please do not hesitate to follow the improvement plan too.

6.6 Recommended Inference Functionality with AMD GPUs

In partnership with the AMD team, we have actually accomplished Day-One assistance for AMD GPUs using SGLang, with complete compatibility for both FP8 and BF16 accuracy. For detailed assistance, please refer to the SGLang instructions.

6.7 Recommended Inference Functionality with Huawei Ascend NPUs

The MindIE framework from the Huawei Ascend neighborhood has effectively adjusted the BF16 version of DeepSeek-V3. For detailed assistance on Ascend NPUs, please follow the guidelines here.

7. License

This is certified under the MIT License. Making use of DeepSeek-V3 Base/Chat designs undergoes the Model License. DeepSeek-V3 series (including Base and Chat) supports industrial usage.