# DeepSeek-V3 / R1 (671B, 37B Active): Shared Expert + Fine-Grained Routing — LoRA Hangi Parçaya?

> Source: https://sukruyusufkaya.com/learn/fine-tuning-cookbook/ftc-deepseek-v3-r1-moe-shared-expert
> Updated: 2026-08-21T17:36:36.635Z
> Category: Fine-Tuning Cookbook (Model-by-Model)
> Module: Part V — MoE Internals & Fine-Tuning
**TLDR:** DeepSeek-V3 (671B param, 37B active) — modern MoE'in en iyi açık örneği. Shared expert (her token'a giden 'common knowledge') + 256 routed expert (fine-grained). DeepSeek-R1 aynı mimari + RL ile reasoning. RTX 4090'da impossible; cookbook'un cloud reçetesi 16×H100 NDR IB + ZeRO-Infinity + expert parallelism.