← sources

Deepseek training 2T and plans 8T model

source: reddit (reddit:LocalLLaMA)published 2026-09-21 12:38ingested 2026-09-22 02:40

retained text

Quote: DeepSeek is training a 2T-parameter model and plans to eventually build an 8T-parameter model.

[https://x.com/wallstengine/status/2101982843656388644](https://x.com/wallstengine/status/2101982843656388644)

Current DeepSeek models:

1. Flash parameter count of **552 billion**
2. **Pro:** **1.6T** (trillion) total parameters with **49B** (billion) activated weights per token

Mythos / Fable is estimated to be 10T parameter count.

document fetched 2026-09-22 00:41 · canonical https://reddit.com/r/LocalLLaMA/comments/1wmblx1/deepseek_training_2t_and_plans_8t_model