Published:
This model is incredible for its size, and it did really well in some coding tasks.
We’re thrilled to announce that DeepSeek V3 03-24 is now available on the Chat-O platform! This latest iteration of the DeepSeek series has shown very strong performance for a compact yet powerful AI model. DeepSeek V3 builds upon its predecessors by delivering faster, smarter, and more efficient performance. Whether you’re writing Python scripts, debugging Elixir code, or exploring creative solutions in JavaScript, this model has your back.
DeepSeek V3 03-24 employs a Mixture-of-Experts (MoE) architecture with 671 billion total parameters, activating only 37 billion per token. This extreme sparsity — less than 6% activation rate — is what made the model so efficient. Key architectural innovations include:
These design choices made DeepSeek V3 one of the most efficient large models ever created. It could run on as little as 4x NVIDIA H800 GPUs in FP8 mode, which was remarkable for a model of its quality class.
| Specification | DeepSeek V3 03-24 |
|---|---|
| Total Parameters | 671B |
| Active Parameters per Token | 37B |
| Architecture | MoE (256 experts, 1 shared) |
| Context Window | 128K tokens |
| Training Tokens | 14.8 trillion |
| Training Hardware | 2,048 H800 GPUs |
| Training Duration | ~2.8 million GPU-hours |
Looking back from 2026, DeepSeek V3 03-24 represents a crucial turning point in the open-weight AI landscape. It was the first model to convincingly demonstrate that MoE architectures with extreme sparsity could match the quality of dense models like GPT-4 while costing a fraction to serve. This insight directly influenced the architecture of subsequent models, including Kimi K2.6 and GLM-4.7, both of which adopted similar sparse MoE designs.
The Kimi K2.7 Code model that followed several months later specifically improved upon DeepSeek V3’s code generation capabilities, achieving a 91.2% pass@1 on HumanEval compared to DeepSeek V3’s already impressive 87.6%.
DeepSeek V3 03-24 achieved scores that stunned the AI community given its training budget:
| Benchmark | DeepSeek V3 | Llama 3.1 405B | GPT-4 (baseline) |
|---|---|---|---|
| MMLU | 88.5% | 87.3% | 86.4% |
| HumanEval pass@1 | 87.6% | 84.1% | 87.0% |
| MATH-500 | 90.2% | 86.5% | 88.9% |
| GSM8K | 92.3% | 89.7% | 92.0% |
| LiveCodeBench | 46.3% | 33.5% | 41.8% |
The LiveCodeBench result was particularly striking — DeepSeek V3 outperformed all existing models on real-world coding challenges that required understanding new libraries and APIs, not just static benchmark problems.
One of the standout features of DeepSeek V3 is its exceptional ability to understand and generate code across multiple programming languages. In our internal testing, the model consistently outperformed expectations in tasks like code completion, bug detection, and even generating complex algorithms. With its lightweight architecture, it delivers these results without compromising on speed or resource usage. For developers who need precision and efficiency, DeepSeek V3 is a dream come true.
The model showed particular strength in:
For developers looking for even more capable coding models in 2026, Kimi K2.7 Code builds directly on DeepSeek V3’s foundation with enhanced training objectives for code tasks. The GLM-5.2 also offers superior coding performance while adding multimodal capabilities that DeepSeek V3 lacked. For general-purpose work, Kimi K3 represents the current state of the art in the open-weight MoE paradigm that DeepSeek V3 pioneered.
At Chat-O, we believe in delivering value through affordability, privacy, and ease of use. Our platform is designed to be cost-effective, ensuring that cutting-edge AI tools like DeepSeek V3 are accessible to everyone. We prioritize your data security, offering a private and secure environment for all your projects. Plus, with intuitive features that make it easy to organize and manage your work, Chat-O is the perfect companion for developers of all skill levels.
Ready to experience the magic of DeepSeek V3 and many other models? Join now and unlock the full potential of AI, with amazing image editing and image generation, strong privacy and much improved performance!
Happy Chat-o-ing!