DeepSeek-V2
#127 en Modèles de langage open sourceDeepSeek · v2 · depuis 2024-05-06 · 2× · vu le 30 juin 2026
1
Momentum
DeepSeek-V2 is a Mixture-of-Experts (MoE) language model by DeepSeek, released in May 2024. It has 236 billion total parameters, of which only 21 billion are activated per token, and supports a 128,000-token context window. The model employs two key architectural innovations: Multi-head Latent Attention (MLA) for KV-cache compression and DeepSeekMoE for efficient sparse computation. Compared to its predecessor DeepSeek 67B, it reduces the KV cache by 93.3% and achieves up to 5.76× the generation throughput.
Historique du momentum
19.05.17.08.
Fonctionnalités
| Benchmark Score (MMLU/Similar) | MMLU (5-shot): 78.5% (DeepSeek-V2 Base); Chat variant: 78.1% MMLU per DeepSeek-Coder-V2 paper |
| Inference Speed | Generation throughput >50,000 tokens/s (on 1 node with 8× H800 GPUs, FP8 precision); prompt input throughput >100,000 tokens/s; equals 5.76× the throughput of DeepSeek 67B |
| Context Window | 128,000 tokens |
| Model Size (Parameters) | 236B total parameters (MoE); 21B activated parameters per token |
| Price Tier | API (at release): approx. $0.14/M input tokens and $0.28/M output tokens; open-weights model available free of charge (DeepSeek License Agreement, commercial use permitted) |
| Memory Requirement | Full model (BF16): at least 8× 80 GB GPUs recommended (e.g., 8× H800/H100); with 4-bit quantization approx. 136 GB VRAM (multi-GPU required) |