Show HN: DeepSeek v3 – A 671B parameter AI Language Model https://ift.tt/XS0hEug

Show HN: DeepSeek v3 – A 671B parameter AI Language Model Hi HN! We're excited to present DeepSeek v3, an advanced language model that combines massive scale with innovative architecture design. Technical Specifications: - 671B total parameters (37B activated per token) - 128K context window - Multi-Token Prediction capability - Auxiliary-loss-free load balancing - FP8 and BF16 inference support Deployment Options: - Multiple framework support (SGLang, LMDeploy, TensorRT-LLM, vLLM) - Hardware compatibility with NVIDIA GPUs, AMD GPUs, Huawei Ascend NPUs - Efficient inference despite large model size - Flexible API integration options Training Details: - 14.8T tokens of diverse, high-quality data - Stable training process with no irrecoverable loss spikes - Efficient cross-node MoE training - Only 2.788M H800 GPU hours for pre-training Benchmark Results: - Superior performance in mathematics and coding tasks - Competitive results against leading closed-source models - Excellent multilingual capabilities - Strong reasoning and understanding abilities Live demo: https://deepseekv3.org/ Looking forward to your feedback on model performance and deployment experience. Future plans include expanding hardware support and optimizing inference efficiency further. https://deepseekv3.org/ December 27, 2024 at 06:27AM