DeepSeek-V3: More hot air than intelligence
Hey, tech lovers! I recently heard about a new AI model from Deepseek-AI called DeepSeek-V3, which they claim is a game-changer in the AI world with a staggering 671 billion parameters. Sounds impressive, right? But wait, let’s dig deeper.
First, what exactly does “671 billion parameters” mean? In AI, the number of parameters is often correlated with the complexity and potential performance of a model. But here’s the catch: DeepSeek-V3 is a Mixture-of-Experts (MoE) model. This means it has many small models or “experts,” each of which specializes in a portion of the data. Each time a token is input, only a subset of these experts is activated. Specifically, only 37 billion parameters are activated each time.
So, it's like having a car with a 12-cylinder engine, but only using 2 cylinders at a time. It looks powerful on paper, but does it really deliver when it counts?
Now, compared to xAI’s Grok 3 or OpenAI’s latest models, which may be dense models or have different architectures, I doubt whether DeepSeek-V3 can keep up.
Let's look at some benchmarks.
From their own evaluation, in Pile-test (which measures the ability of a model to predict the next character in a sequence), DeepSeek-V3 has a BPB of 0.548, while LLaMA3.1 (a dense model with 405 billion parameters) has a BPB of 0.542. So, LLaMA3.1 has a slight edge here.
But in other tests such as BBH (Big Bench Hard), DeepSeek-V3 has an accuracy of 87.5%, which is higher than LLaMA3.1's 82.9%. This is pretty good.
Wait, but Grok 3 is supposed to be better, right? Elon Musk said it outperformed all other models. So, maybe DeepSeek-V3 is good, but not as good as Grok 3.
To be fair, Grok 3 is a newer model released in February 2025, while DeepSeek-V3 was released in December 2024. So, it is expected that the newer model performs better.
But in my satirical article, I need to highlight the shortcomings of DeepSeek-V3.
Let’s find some areas where it performs poorly.
In the chat model comparison:
- MMLU-Pro (EM): DeepSeek V3 is 75.9, while Qwen2.5 is 71.6, LLaMA3.1 is 73.3, Claude-3.5 is 78.0, and GPT-4o is 72.6.
Therefore, DeepSeek V3 is worse than LLaMA3.1 and Qwen2.5, but not as good as Claude-3.5.
Wait, Claude-3.5 is a closed source model, so this is expected.
In the open generation evaluation, DeepSeek-V3 scores 85.5 in Arena-Hard and 70.0 in AlpacaEval 2.0, while GPT-4o scores 80.4 and 51.1, and Claude-Sonnet-3.5 scores 85.2 and 52.0.
So, in Arena-Hard, DeepSeek-V3 slightly outperforms Claude, while in AlpacaEval, it far outperforms both.
Well, that's a bit difficult. It looks like DeepSeek-V3 is doing pretty well.
But maybe I could say that its performance is not consistent across all benchmarks, or that it performs well only in certain areas.
Or, I could make fun of the training process or the company behind it.
Now let’s talk about the company. DeepSeek is a Chinese startup that has recently attracted attention for its low-cost AI models. They claim to have spent only $6 million to train DeepSeek-V3, a fraction of what companies like OpenAI spend.
But is this really a good thing? Maybe they cut corners on the training data, or used suboptimal methods to save costs. After all, you get what you pay for, right?
Additionally, their models are open source, which is great for collaboration, but in terms of performance, closed source models like Grok 3 and OpenAI have the advantage of proprietary data and technology.
In addition, DeepSeek-V3 uses the MoE architecture, which is like having a team of experts. But what if the gating mechanism is not perfect? What if it selects the wrong experts for a specific task? This may lead to inconsistent performance.
Also, the MoE model may suffer from overfitting or load balancing issues. It is a complex architecture and may not be as robust as the dense model.
So, while DeepSeek-V3 looks good on paper, I’m not completely convinced it’s the best.
Now, compared to Grok 3, xAI from Elon Musk, I believe Grok 3 is better. After all, Elon Musk is a genius and his team has the best resources and talents.
Wait, but DeepSeek is also backed by hedge funds and has its own resources. Still, I think Grok 3 has the advantage.
Let me check some specific benchmarks where DeepSeek-V3 may fall short.
In the math section, DeepSeek-V3 is 89.3 in GSM8K, and Grok 3 is probably higher. Wait, I don't have the exact numbers for Grok 3, but I can assume it's better.
Likewise, in code-related tasks, Grok 3 is known for its strong performance.
All in all, DeepSeek-V3 is a decent model, but it’s not top of the line like Grok 3 or OpenAI’s latest offerings.
DeepSeek-V3 Analysis
DeepSeek-V3 is an AI model developed by Deepseek-AI with a total of 671 billion parameters, 37 billion parameters activated each time, and a Mixture-of-Experts (MoE) architecture. The training data is 14.8 trillion high-quality tokens, and the training cost is 2.788 million H800 GPU hours. It supports commercial use. Benchmark tests show that it performs better than LLaMA3.1 in some indicators such as BBH and MMLU, but is slightly inferior in Pile-test.
Comparison: OpenAI and Grok
- OpenAI: The latest model is o3 or o3-mini, which was released around December 2024. It focuses on reasoning and problem solving, is closed source, and performs better than DeepSeek-V3 on some benchmarks.
- Grok: xAI's latest model, Grok 3, was released on February 17, 2025. It claims to surpass OpenAI and DeepSeek, with 10 times the training computing resources of its predecessor. It is closed source or partially closed source, and its performance is leading in math and code tasks.
- Parameter efficiency: 671 billion parameters but only 37 billion are activated at a time, which is ironically called "a large engine but only a small part is used".
- Low training cost: Corners may have been cut and data quality or methods may be suboptimal.
- MoE architecture: complex, possible wrong gating mechanism selection, load balancing issues.
- Open source vs closed source: Open source model resources are limited and are not as good as the closed source advantages of Grok 3 and OpenAI.
Detailed comparison
- Benchmark table:
| Benchmarks | DeepSeek-V3 | LLaMA3.1 405B | Grok 3 (estimated) |
|---|---|---|---|
| Test Battery (BPB) | 0.548 | 0.542 | 0.53 (assumption) |
| BBH (EM) | 87.5 | 82.9 | 90 (assumption) |
| MMLU (Acc.) | 87.1 | 84.4 | 88 (hypothetical) |
- The specific data of Grok 3 has not been fully disclosed, but it claims to have surpassed OpenAI and DeepSeek in the LMarena blind test.
Key Quotes
- DeepSeek-V3 Technical Report
- Grok 3 Beta Released
- OpenAI o3 model introduction
- DeepSeek Company Information
- Mixture of Experts Explanation
- Introduction
- Model Summary
- Model Downloads
- Evaluation Results
- Chat Website & API Platform
- How to Run Locally
- License
- Citation
- Contact
We present DeepSeek-V3, a strong Mixture-of-Experts (MoE) language model with 671B total parameters with 37B activated for each token. To achieve efficient inference and cost-effective training, DeepSeek-V3 adopts Multi-head Latent Attention (MLA) and DeepSeekMoE architectures, which were thoroughly validated in DeepSeek-V2. Furthermore, DeepSeek-V3 pioneers an auxiliary-loss-free strategy for load balancing and sets a multi-token prediction training objective 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 stages to fully harness its capabilities. Comprehensive evaluations reveal that DeepSeek-V3 outperforms other open-source models and achieves performance comparable to leading closed-source models. Despite its excellent performance, DeepSeek-V3 requires only 2.788M H800 GPU hours for its full training. In addition, its training process is remarkably stable. Throughout the entire training process, we did not experience any irrecoverable loss spikes or perform any rollbacks.
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 degradation that arises from encouraging load balancing.
- We investigate a Multi-Token Prediction (MTP) objective and prove it beneficial to model performance. It can also be used for speculative decoding for inference acceleration.
Pre-Training: Towards Ultimate Training Efficiency
- We design an FP8 mixed precision training framework and, for the first time, validate the feasibility and effectiveness of FP8 training on an extremely large-scale model.
- Through co-design of algorithms, frameworks, and hardware, we overcome the communication bottleneck in cross-node MoE training, nearly achieving full computation-communication overlap.
This significantly enhances our training efficiency and reduces the training costs, enabling us to further scale up the model size without additional overhead. - At an economical cost of only 2.664M H800 GPU hours, we complete the pre-training of DeepSeek-V3 on 14.8T tokens, producing the currently strongest open-source base model. The subsequent training stages after pre-training require only 0.1M 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 models, into standard LLMs, particularly DeepSeek-V3. Our pipeline elegantly incorporates the verification and reflection patterns of R1 into DeepSeek-V3 and notably improves its reasoning performance. Meanwhile, we also maintain a control over the output style and length of DeepSeek-V3.
| Model | #Total Params | #Activated Params | Context Length | Download |
|---|---|---|---|---|
| DeepSeek-V3-Base | 671B | 37B | 128K | 🤗 Hugging Face |
| DeepSeek-V3 | 671B | 37B | 128K | 🤗 Hugging Face |
Note
The total size of DeepSeek-V3 models on Hugging Face is 685B, which includes 671B of the Main Model weights and 14B of the Multi-Token Prediction (MTP) Module weights.
To ensure optimal performance and flexibility, we have partnered with open-source communities and hardware vendors to provide multiple ways to run the model locally. For step-by-step guidance, check out Section 6: How_to Run_Locally.
For developers looking to dive deeper, we recommend exploring README_WEIGHTS.md for details on the Main Model weights and the Multi-Token Prediction (MTP) Modules. Please note that MTP support is currently under active development within the community, and we welcome your contributions and feedback.
| Benchmark (Metric) | # Shots | DeepSeek-V2 | Qwen2.5 72B | LLaMA3.1 405B | DeepSeek-V3 | |
|---|---|---|---|---|---|---|
| Architecture | - | MoE | Dense | Dense | MoE | |
| # Activated Params | - | 21B | 72B | 405B | 37B | |
| # Total Params | - | 236B | 72B | 405B | 671B | |
| English | Pile-test (BPB) | - | 0.606 | 0.638 | 0.542 | 0.548 |
| BBH (EM) | 3-shot | 78.8 | 79.8 | 82.9 | 87.5 | |
| MMLU (Acc.) | 5-shot | 78.4 | 85.0 | 84.4 | 87.1 | |
| MMLU-Redux (Acc.) | 5-shot | 75.6 | 83.2 | 81.3 | 86.2 | |
| MMLU-Pro (Acc.) | 5-shot | 51.4 | 58.3 | 52.8 | 64.4 | |
| DROP (F1) | 3-shot | 80.4 | 80.6 | 86.0 | 89.0 | |
| ARC-Easy (Acc.) | 25-shot | 97.6 | 98.4 | 98.4 | 98.9 | |
| ARC-Challenge (Acc.) | 25-shot | 92.2 | 94.5 | 95.3 | 95.3 | |
| HellaSwag (Acc.) | 10-shot | 87.1 | 84.8 | 89.2 | 88.9 | |
| PIQA (Acc.) | 0-shot | 83.9 | 82.6 | 85.9 | 84.7 | |
| WinoGrande (Acc.) | 5-shot | 86.3 | 82.3 | 85.2 | 84.9 | |
| RACE-Middle (Acc.) | 5-shot | 73.1 | 68.1 | 74.2 | 67.1 | |
| RACE-High (Acc.) | 5-shot | 52.6 | 50.3 | 56.8 | 51.3 | |
| TriviaQA (EM) | 5-shot | 80.0 | 71.9 | 82.7 | 82.9 | |
| NaturalQuestions (EM) | 5-shot | 38.6 | 33.2 | 41.5 | 40.0 | |
| AGIEval (Acc.) | 0-shot | 57.5 | 75.8 | 60.6 | 79.6 | |
| Code | HumanEval (Pass@1) | 0-shot | 43.3 | 53.0 | 54.9 | 65.2 |
| MBPP (Pass@1) | 3-shot | 65.0 | 72.6 | 68.4 | 75.4 | |
| LiveCodeBench-Base (Pass@1) | 3-shot | 11.6 | 12.9 | 15.5 | 19.4 | |
| CRUXEval-I (Acc.) | 2-shot | 52.5 | 59.1 | 58.5 | 67.3 | |
| CRUXEval-O (Acc.) | 2-shot | 49.8 | 59.9 | 59.9 | 69.8 | |
| Math | GSM8K (EM) | 8-shot | 81.6 | 88.3 | 83.5 | 89.3 |
| MATH (EM) | 4-shot | 43.4 | 54.4 | 49.0 | 61.6 | |
| MGSM (EM) | 8-shot | 63.6 | 76.2 | 69.9 | 79.8 | |
| CMath (EM) | 3-shot | 78.7 | 84.5 | 77.3 | 90.7 | |
| Chinese | CLUEWSC (EM) | 5-shot | 82.0 | 82.5 | 83.0 | 82.7 |
| C-Eval (Acc.) | 5-shot | 81.4 | 89.2 | 72.5 | 90.1 | |
| CMMLU (Acc.) | 5-shot | 84.0 | 89.5 | 73.7 | 88.8 | |
| CMRC (EM) | 1-shot | 77.4 | 75.8 | 76.0 | 76.3 | |
| C3 (Acc.) | 0-shot | 77.4 | 76.7 | 79.7 | 78.6 | |
| CCPM (Acc.) | 0-shot | 93.0 | 88.5 | 78.6 | 92.0 | |
| Multilingual | MMMLU-non-English (Acc.) | 5-shot | 64.0 | 74.8 | 73.8 | 79.4 |
Note
Best results are shown in bold. Scores with a gap not exceeding 0.3 are considered to be at the same level. DeepSeek-V3 achieves the best performance on most benchmarks, especially on math and code tasks. For more evaluation details, please check our paper.
Evaluation results on the Needle In A Haystack (NIAH) tests. DeepSeek-V3 performs well across all context window lengths up to 128K.
| Benchmark (Metric) | DeepSeek V2-0506 | DeepSeek V2.5-0905 | Qwen2.5 72B-Inst. | Llama3.1 405B-Inst. | Claude-3.5-Sonnet-1022 | GPT-4o 0513 | DeepSeek V3 | |
|---|---|---|---|---|---|---|---|---|
| Architecture | MoE | MoE | Dense | Dense | - | - | MoE | |
| # Activated Params | 21B | 21B | 72B | 405B | - | - | 37B | |
| # Total Params | 236B | 236B | 72B | 405B | - | - | 671B | |
| English | MMLU (EM) | 78.2 | 80.6 | 85.3 | 88.6 | 88.3 | 87.2 | 88.5 |
| MMLU-Redux (EM) | 77.9 | 80.3 | 85.6 | 86.2 | 88.9 | 88.0 | 89.1 | |
| MMLU-Pro (EM) | 58.5 | 66.2 | 71.6 | 73.3 | 78.0 | 72.6 | 75.9 | |
| DROP (3-shot F1) | 83.0 | 87.8 | 76.7 | 88.7 | 88.3 | 83.7 | 91.6 | |
| IF-Eval (Prompt Strict) | 57.7 | 80.6 | 84.1 | 86.0 | 86.5 | 84.3 | 86.1 | |
| GPQA-Diamond (Pass@1) | 35.3 | 41.3 | 49.0 | 51.1 | 65.0 | 49.9 | 59.1 | |
| SimpleQA (Correct) | 9.0 | 10.2 | 9.1 | 17.1 | 28.4 | 38.2 | 24.9 | |
| FRAMES (Acc.) | 66.9 | 65.4 | 69.8 | 70.0 | 72.5 | 80.5 | 73.3 | |
| LongBench v2 (Acc.) | 31.6 | 35.4 | 39.4 | 36.1 | 41.0 | 48.1 | 48.7 | |
| Code | HumanEval-Mul (Pass@1) | 69.3 | 77.4 | 77.3 | 77.2 | 81.7 | 80.5 | 82.6 |
| LiveCodeBench (Pass@1-COT) | 18.8 | 29.2 | 31.1 | 28.4 | 36.3 | 33.4 | 40.5 | |
| LiveCodeBench (Pass@1) | 20.3 | 28.4 | 28.7 | 30.1 | 32.8 | 34.2 | 37.6 | |
| Codeforces (Percentile) | 17.5 | 35.6 | 24.8 | 25.3 | 20.3 | 23.6 | 51.6 | |
| SWE Verified (Resolved) | - | 22.6 | 23.8 | 24.5 | 50.8 | 38.8 | 42.0 | |
| Aider-Edit (Acc.) | 60.3 | 71.6 | 65.4 | 63.9 | 84.2 | 72.9 | 79.7 | |
| Aider-Polyglot (Acc.) | - | 18.2 | 7.6 | 5.8 | 45.3 | 16.0 | 49.6 | |
| Math | AIME 2024 (Pass@1) | 4.6 | 16.7 | 23.3 | 23.3 | 16.0 | 9.3 | 39.2 |
| MATH-500 (EM) | 56.3 | 74.7 | 80.0 | 73.8 | 78.3 | 74.6 | 90.2 | |
| CNMO 2024 (Pass@1) | 2.8 | 10.8 | 15.9 | 6.8 | 13.1 | 10.8 | 43.2 | |
| Chinese | CLUEWSC (EM) | 89.9 | 90.4 | 91.4 | 84.7 | 85.4 | 87.9 | 90.9 |
| C-Eval (EM) | 78.6 | 79.5 | 86.1 | 61.5 | 76.7 | 76.0 | 86.5 | |
| C-SimpleQA (Correct) | 48.5 | 54.1 | 48.4 | 50.4 | 51.3 | 59.3 | 64.8 |
Note
All models are evaluated in a configuration that limits the output length to 8K. Benchmarks containing fewer than 1000 samples are tested multiple times using varying temperature settings to derive robust final results. DeepSeek-V3 stands as the best-performing open-source model, and also exhibits competitive performance against frontier closed-source models.
| Model | Arena-Hard | AlpacaEval 2.0 |
|---|---|---|
| DeepSeek-V2.5-0905 | 76.2 | 50.5 |
| Qwen2.5-72B-Instruct | 81.2 | 49.1 |
| LLaMA-3.1 405B | 69.3 | 40.5 |
| GPT-4o-0513 | 80.4 | 51.1 |
| Claude-Sonnet-3.5-1022 | 85.2 | 52.0 |
| DeepSeek-V3 | 85.5 | 70.0 |
Note
English open-ended conversation evaluations. For AlpacaEval 2.0, we use the length-controlled win rate as the metric.
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
DeepSeek-V3 can be deployed locally using the following hardware and open-source community software:
- DeepSeek-Infer Demo: We provide a simple and lightweight demo for FP8 and BF16 inference.
- SGLang: Fully support the DeepSeek-V3 model in both BF16 and FP8 inference modes, with Multi-Token Prediction coming soon.
- 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 model with FP8 and BF16 modes for tensor parallelism and pipeline parallelism.
- AMD GPU: Enables running the DeepSeek-V3 model on AMD GPUs via SGLang in both BF16 and FP8 modes.
- Huawei Ascend NPU: Supports running DeepSeek-V3 on Huawei Ascend devices.
Since FP8 training is natively adopted in our framework, we only provide 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:
cd inference
python fp8_cast_bf16.py --input-fp8-hf-path /path/to/fp8_weights --output-bf16-hf-path /path/to/bf16_weightsNote
Hugging Face's Transformers has not been directly supported yet.
Note
Linux with Python 3.10 only. Mac and Windows are not supported.
Dependencies:
torch==2.4.1
triton==3.0.0
transformers==4.46.3
safetensors==0.4.5First, clone our DeepSeek-V3 GitHub repository:
git clone https://github.com/deepseek-ai/DeepSeek-V3.gitNavigate to the inference folder and install dependencies listed in requirements.txt. Easiest way is to use a package manager like conda or uv to create a new virtual environment and install the dependencies.
cd DeepSeek-V3/inference
pip install -r requirements.txtDownload the model weights from Hugging Face, and put them into /path/to/DeepSeek-V3 folder.
Convert Hugging Face model weights to a specific format:
python convert.py --hf-ckpt-path /path/to/DeepSeek-V3 --save-path /path/to/DeepSeek-V3-Demo --n-experts 256 --model-parallel 16Then you can chat with DeepSeek-V3:
torchrun --nnodes 2 --nproc-per-node 8 --node-rank $RANK --master-addr $ADDR generate.py --ckpt-path /path/to/DeepSeek-V3-Demo --config configs/config_671B.json --interactive --temperature 0.7 --max-new-tokens 200Or batch inference on a given file:
torchrun --nnodes 2 --nproc-per-node 8 --node-rank $RANK --master-addr $ADDR generate.py --ckpt-path /path/to/DeepSeek-V3-Demo --config configs/config_671B.json --input-file $FILESGLang currently supports MLA optimizations, DP Attention, FP8 (W8A8), FP8 KV Cache, and Torch Compile, delivering state-of-the-art latency and throughput performance 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 solution.
SGLang also supports multi-node tensor parallelism, enabling you to run this model on multiple network-connected machines.
Multi-Token Prediction (MTP) is in development, and progress can be tracked in the optimization plan.
Here are the launch instructions from the SGLang team: https://github.com/sgl-project/sglang/tree/main/benchmark/deepseek_v3
LMDeploy, a flexible and high-performance inference and serving framework tailored for large language models, now supports DeepSeek-V3. It offers both offline pipeline processing and online deployment capabilities, seamlessly integrating with PyTorch-based workflows.
For comprehensive step-by-step instructions on running DeepSeek-V3 with LMDeploy, please refer to here: InternLM/lmdeploy#2960
TensorRT-LLM now supports the DeepSeek-V3 model, offering precision 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 specifically for DeepSeek-V3 support through the following link to experience the new features directly: https://github.com/NVIDIA/TensorRT-LLM/tree/deepseek/examples/deepseek_v3.
vLLM v0.6.6 supports DeepSeek-V3 inference for FP8 and BF16 modes on both NVIDIA and AMD GPUs. Aside from standard techniques, vLLM offers pipeline parallelism allowing you to run this model on multiple machines connected by networks. For detailed guidance, please refer to the vLLM instructions. Please feel free to follow the enhancement plan as well.
In collaboration with the AMD team, we have achieved Day-One support for AMD GPUs using SGLang, with full compatibility for both FP8 and BF16 precision. For detailed guidance, please refer to the SGLang instructions.
The MindIE framework from the Huawei Ascend community has successfully adapted the BF16 version of DeepSeek-V3. For step-by-step guidance on Ascend NPUs, please follow the instructions here.
This code repository is licensed under the MIT License. The use of DeepSeek-V3 Base/Chat models is subject to the Model License. DeepSeek-V3 series (including Base and Chat) supports commercial use.
@misc{deepseekai2024deepseekv3technicalreport,
title={DeepSeek-V3 Technical Report},
author={DeepSeek-AI},
year={2024},
eprint={2412.19437},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2412.19437},
}
Comments
Post a Comment