GitHub – Deepseek-ai/DeepSeek-V3

We present DeepSeek-V3, a strong Mixture-of-Experts (MoE) language design with 671B overall specifications with 37B triggered for each token. To achieve efficient reasoning and cost-efficient training, DeepSeek-V3 embraces Multi-head Latent Attention (MLA) and DeepSeekMoE architectures, which were thoroughly validated 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 diverse and high-quality tokens, followed by Supervised Fine-Tuning and Reinforcement Learning stages to fully harness its capabilities. Comprehensive examinations expose that DeepSeek-V3 exceeds other open-source designs and attains efficiency comparable to leading closed-source models. Despite its outstanding performance, DeepSeek-V3 needs only 2.788 M H800 GPU hours for its full training. In addition, its training process is incredibly steady. Throughout the whole training process, 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 pioneer an auxiliary-loss-free method for load balancing, which minimizes the efficiency deterioration that occurs from motivating load balancing.
– We examine a (MTP) objective and show it advantageous to model efficiency. It can also be used for speculative decoding for reasoning velocity.

Pre-Training: Towards Ultimate Training Efficiency

– We create an FP8 combined accuracy training framework and, for the very first time, verify the expediency and effectiveness of FP8 training on an extremely large-scale model.
– Through co-design of algorithms, structures, and hardware, we overcome the communication bottleneck in cross-node MoE training, almost accomplishing full computation-communication overlap.
This substantially improves our training efficiency and lowers the training expenses, allowing us to further scale up the design size without additional overhead.
– At an affordable 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 phases after pre-training need only 0.1 M GPU hours.

Post-Training: Knowledge Distillation from DeepSeek-R1

– We introduce an innovative methodology to distill reasoning capabilities from the long-Chain-of-Thought (CoT) model, specifically from one of the DeepSeek R1 series designs, into standard LLMs, particularly DeepSeek-V3. Our pipeline elegantly incorporates the verification and reflection patterns of R1 into DeepSeek-V3 and significantly enhances its thinking performance. Meanwhile, we also keep a control over the output design 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 ensure optimal performance and versatility, we have partnered with open-source communities and hardware vendors to provide several methods to run the design locally. For detailed guidance, examine out Section 6: How_to Run_Locally.

For developers seeking to dive much deeper, we advise checking out README_WEIGHTS. md for information on the Main Model weights and the Multi-Token Prediction (MTP) Modules. Please note that MTP support 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 strong. Scores with a gap not going beyond 0.3 are considered to be at the very same level. DeepSeek-V3 accomplishes the finest performance on many criteria, particularly on math and code tasks. For more assessment information, 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 approximately 128K.

Chat Model

Standard Benchmarks (Models larger than 67B)

All designs are evaluated in a configuration that limits the output length to 8K. Benchmarks including fewer than 1000 samples are tested several times using varying temperature level settings to derive robust results. DeepSeek-V3 stands as the best-performing open-source model, and also displays competitive performance versus frontier closed-source designs.

Open Ended Generation Evaluation

English open-ended conversation examinations. 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 site: chat.deepseek.com

We likewise offer OpenAI-Compatible API at DeepSeek Platform: platform.deepseek.com

6. How to Run Locally

DeepSeek-V3 can be released in your area utilizing the following hardware and open-source community software:

DeepSeek-Infer Demo: We supply a simple and lightweight demonstration for FP8 and BF16 reasoning.
SGLang: Fully support the DeepSeek-V3 model in both BF16 and FP8 reasoning modes, with Multi-Token Prediction coming quickly.
LMDeploy: Enables effective FP8 and BF16 inference for local and cloud release.
TensorRT-LLM: Currently supports BF16 reasoning and INT4/8 quantization, with FP8 assistance 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 through SGLang in both BF16 and FP8 modes.
Huawei Ascend NPU: Supports running DeepSeek-V3 on Huawei Ascend gadgets.
Since FP8 training is natively adopted in our structure, we only offer FP8 weights. If you need BF16 weights for experimentation, you can utilize the provided conversion script to perform the change.

Here is an example of transforming 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 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 noted in requirements.txt. Easiest way is to use a package supervisor like conda or uv to develop a brand-new virtual environment and install the dependences.

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 inference on a provided file:

6.2 Inference with SGLang (recommended)

SGLang presently supports MLA optimizations, DP Attention, FP8 (W8A8), FP8 KV Cache, and Torch Compile, delivering advanced latency and throughput performance amongst open-source structures.

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

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

Multi-Token Prediction (MTP) remains in advancement, and progress can be tracked in the optimization strategy.

Here are the launch instructions 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 structure customized for big language models, now supports DeepSeek-V3. It offers both offline pipeline processing and online deployment 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 (recommended)

TensorRT-LLM now supports the DeepSeek-V3 model, providing accuracy choices such as BF16 and INT4/INT8 weight-only. Support for FP8 is currently in progress and will be released quickly. You can access the customized branch of TRTLLM specifically 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 (suggested)

vLLM v0.6.6 supports DeepSeek-V3 inference for FP8 and BF16 modes on both NVIDIA and AMD GPUs. Aside from standard strategies, vLLM offers pipeline parallelism allowing you to run this model on multiple machines linked by networks. For comprehensive guidance, please refer to the vLLM guidelines. Please feel totally free to follow the enhancement strategy also.

6.6 Recommended Inference Functionality with AMD GPUs

In collaboration with the AMD group, we have accomplished Day-One support for AMD GPUs using SGLang, with complete compatibility for both FP8 and BF16 precision. For in-depth guidance, please refer to the SGLang instructions.

6.7 Recommended Inference Functionality with Huawei Ascend NPUs

The MindIE framework from the Huawei Ascend neighborhood has actually successfully adjusted the BF16 variation of DeepSeek-V3. For step-by-step guidance on Ascend NPUs, please follow the directions here.

7. License

This code repository is accredited under the MIT License. Using DeepSeek-V3 Base/Chat models undergoes the Model License. DeepSeek-V3 series (including Base and Chat) supports commercial usage.