chiprook
← AI
AIOctober 4, 2026, 00:00

DeepSeek 4.1: 670B parameters, 256 experts and MTP-4x

DeepSeek-AI released open weights and a corporate endpoint for DeepSeek 4.1: 670B total parameters with only 37B active per token, 256 routed experts plus one shared expert. MTP-4x and DualPipe 2.0 push throughput past 240 tokens/s per node at $0.55 per million output tokens.

DeepSeek 4.1: 670B parameters, 256 experts and MTP-4x
#DeepSeek
Read next
AI

Firm rents 4x Nvidia H200s, finds DeepSeek '80x cheaper' claim fails

AI

DeepSeek Cuts AI Agent Memory Cost 4x: New Architecture Fits More Sessions Per GPU

Security

ZoomEye Scan Finds 4.1 Million Hosts Exposing Port 6443

AI

Aleph Alpha launches Kolibri, a German 78.1B-parameter MoE model