GitHub – Deepseek-ai/DeepSeek-V3

We present DeepSeek-V3, a strong Mixture-of-Experts (MoE) language design with 671B total specifications with 37B activated for each token. To achieve efficient inference and affordable training, DeepSeek-V3 adopts Multi-head Latent Attention (MLA) and DeepSeekMoE architectures, which were thoroughly verified in DeepSeek-V2. Furthermore, DeepSeek-V3 leaders an auxiliary-loss-free technique for load balancing and sets a multi-token forecast training objective for more powerful performance. We pre-train DeepSeek-V3 on 14.8 trillion diverse and top quality tokens, followed by Supervised Fine-Tuning and Reinforcement Learning stages to fully harness its abilities. Comprehensive assessments expose that DeepSeek-V3 outshines other open-source models and achieves efficiency similar to leading closed-source models. Despite its excellent performance, DeepSeek-V3 needs just 2.788 M H800 GPU hours for its complete training. In addition, its training procedure is extremely 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 effective architecture of DeepSeek-V2, we leader an auxiliary-loss-free technique for load balancing, which decreases the performance deterioration that arises from motivating load balancing.
– We investigate a Multi-Token Prediction (MTP) goal and show it beneficial to design performance. It can also be used for speculative decoding for inference velocity.

Pre-Training: Towards Ultimate Training Efficiency

– We design an FP8 mixed precision training framework and, for the first time, verify the expediency and efficiency of FP8 training on an extremely massive model.
– Through co-design of algorithms, frameworks, and hardware, we get rid of the communication bottleneck in cross-node MoE training, almost achieving full computation-communication overlap.
This substantially improves our training performance and reduces the training costs, allowing us to further scale up the design size without extra overhead.
– At an affordable expense of only 2.664 M H800 GPU hours, we finish the pre-training of DeepSeek-V3 on 14.8 T tokens, producing the presently greatest 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 methodology to boil down reasoning capabilities from the long-Chain-of-Thought (CoT) model, specifically from among the DeepSeek R1 series designs, into basic LLMs, especially DeepSeek-V3. Our pipeline elegantly includes the verification and reflection patterns of R1 into DeepSeek-V3 and notably improves its reasoning efficiency. Meanwhile, we also preserve 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 ensure optimum efficiency and flexibility, we have actually partnered with open-source neighborhoods and hardware vendors to provide multiple ways to run the design in your area. For detailed assistance, take a look at Section 6: How_to Run_Locally.

For designers aiming to dive deeper, we suggest 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 welcome your contributions and feedback.

4. Evaluation Results

Base Model

Standard Benchmarks

Best outcomes are revealed in bold. Scores with a gap not going beyond 0.3 are thought about to be at the same level. DeepSeek-V3 accomplishes the finest efficiency on many standards, particularly on math and code jobs. For more assessment information, please examine our paper.

Context Window

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

Chat Model

Standard Benchmarks (Models larger than 67B)

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

Open Ended Generation Evaluation

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

5. Chat Website & API Platform

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

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

6. How to Run Locally

DeepSeek-V3 can be released locally utilizing the following hardware and open-source community software application:

DeepSeek-Infer Demo: We supply an easy and lightweight demo for FP8 and BF16 inference.
SGLang: Fully support the DeepSeek-V3 design in both BF16 and FP8 inference modes, with Multi-Token Prediction coming quickly.
LMDeploy: Enables efficient FP8 and BF16 reasoning for regional and cloud deployment.
TensorRT-LLM: Currently supports BF16 inference and INT4/8 quantization, with FP8 coming soon.
vLLM: Support DeepSeek-V3 model with FP8 and BF16 modes for tensor parallelism and pipeline parallelism.
AMD GPU: Enables running the DeepSeek-V3 model 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 provide FP8 weights. If you need BF16 weights for experimentation, you can use the supplied conversion script to perform the improvement.

Here is an example of converting FP8 weights to BF16:

Hugging Face’s Transformers has not been straight 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 inference folder and set up dependences noted in requirements.txt. Easiest way is to use a package manager like conda or uv to create a new virtual environment and install the reliances.

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 specific format:

Run

Then you can chat with DeepSeek-V3:

Or batch inference on an offered file:

6.2 Inference with SGLang (advised)

SGLang currently supports MLA optimizations, DP Attention, FP8 (W8A8), FP8 KV Cache, and Torch Compile, providing advanced 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 option.

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

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

Here are the launch directions 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 tailored for large language designs, now supports DeepSeek-V3. It uses both offline pipeline processing and online release abilities, perfectly integrating with PyTorch-based workflows.

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

6.4 Inference with TRT-LLM (suggested)

TensorRT-LLM now supports the DeepSeek-V3 design, using precision alternatives such as BF16 and INT4/INT8 weight-only. Support for FP8 is presently in development and will be launched soon. You can access the customized branch of TRTLLM specifically for DeepSeek-V3 assistance through the following link to experience the new functions straight: https://github.com/NVIDIA/TensorRT-LLM/tree/deepseek/examples/deepseek_v3.

6.5 Inference with vLLM (advised)

vLLM v0.6.6 supports DeepSeek-V3 inference for FP8 and BF16 modes on both NVIDIA and AMD GPUs. Aside from standard methods, vLLM uses pipeline parallelism enabling you to run this design on numerous makers connected by networks. For in-depth assistance, please refer to the vLLM instructions. Please do not hesitate to follow the improvement strategy also.

6.6 Recommended Inference Functionality with AMD GPUs

In partnership with the AMD group, we have achieved Day-One assistance for AMD GPUs utilizing SGLang, with complete compatibility for both FP8 and BF16 accuracy. For in-depth assistance, please describe the SGLang instructions.

6.7 Recommended Inference Functionality with Huawei Ascend NPUs

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

7. License

This code repository is licensed under the MIT License. The usage of DeepSeek-V3 Base/Chat models is subject to the Model License. DeepSeek-V3 series (consisting of Base and Chat) supports industrial usage.