DeepSeek V3 0324 is DeepSeek's open-source 671B Mixture-of-Experts language model released December 26, 2024. It achieves 3x the inference throughput of DeepSeek-V2 while matching closed-source models in published benchmark evaluations.
import { streamText } from 'ai'
const result = streamText({ model: 'deepseek/deepseek-v3', prompt: 'Why is the sky blue?'})DeepSeek V3 0324 was released December 26, 2024 as the third major iteration of DeepSeek's general-purpose language model line. It uses a sparse Mixture-of-Experts (MoE) architecture with 671B total parameters and 37B active per forward pass. Inference compute scales with the 37B active parameters rather than the full 671B, so throughput is roughly 3x that of DeepSeek-V2. Live throughput metrics appear on this page.
The context window of 163.8K tokens was validated through Needle In A Haystack (NIAH) evaluations across the full range, confirming reliable retrieval rather than nominal coverage. The research paper and model weights are openly published.
DeepSeek V3 0324 maintains full API backward compatibility with earlier DeepSeek integrations, reducing migration effort for teams upgrading from V2. Performance evaluations place it on par with closed-source models across knowledge, reasoning, code, and language understanding benchmarks.