DeepSeek 버전 히스토리 | V1, V2, V3, R1, V4 진화 타임라인

First open-source version, offering 7B and 67B scales. 67B version surpasses LLaMA-2 70B in code, math, reasoning tasks. Trained on 2T tokens, proving strength of Chinese team in large models.

Open-source multimodal model, supports 1024×1024 high-resolution image understanding. Excellent performance in multiple vision-language tasks, adding multimodal capability to DeepSeek ecosystem.

Adopts Mixture-of-Experts (MoE) architecture, 236B total params, 21B active, supports 128K context. Training cost reduced 42.5%, KV cache reduced 93.3%, throughput improved 5.76x.

Code-focused MoE model, supports 338 programming languages, 128K context. Additional 6T tokens code data training, HumanEval score 89.5%.

DeepSeek's strongest model, 671B total params, 37B active. Trained on 14.8T tokens, only 2.788M H800 GPU hours needed. Stable training with no rollbacks.

Model focused on complex reasoning, excels in math, programming, logical reasoning tasks.

Released on April 24, 2026 and open-sourced under MIT, with weights on Hugging Face. Two versions: V4-Pro (1.6T total / 49B active) and V4-Flash (284B / 13B). MoE with CSA+HCA hybrid attention delivers a 1M-token context at very low cost (~27% compute, ~10% KV cache per token vs V3.2). SWE-bench Verified 80.6%.