Mmastodon TechnologySemiconductors first seen 7 d ago, last 6 d ago, peak #1

DeepSeek open-sources DeepInfer, revealing massive inference infrastructure

Original: DeepSeek just upended the narrative that Chinese labs keep their infrastructure secret. They've open-sourced DeepInfer,

DeepSeek has released DeepInfer as open source, detailing how its AI systems handle an estimated 155 trillion tokens per day. The release explains the company's architecture, which splits prefill and decode workloads across GPU nodes and relies on a KV cache of around 100 PiB. Observers say the move challenges the assumption that Chinese AI labs keep their infrastructure details hidden, and it is being closely examined by engineers studying large-scale inference efficiency.

Why now: DeepSeek's unusually detailed disclosure of its inference infrastructure contradicts expectations about Chinese AI lab secrecy and offers rare technical insight.

DeepSeekDeepInfer

Open on mastodon →

Evidence

• DeepSeek just upended the narrative that Chinese labs keep their infrastructure secret. They've open-sourced DeepInfer, revealing how they process 155 trillion tokens daily by splitting prefill and decode tasks across GPU nodes and backed by a massive 100 PiB KV cache. A new… · asiaai · 2

API: https://socialmediatrends-api.osmike.com/v1/trends/3670685