by Amit Shekhar · 24 April 2026
Decoding DeepSeek-V4
DeepSeek-V4, the new family of open Mixture-of-Experts language models that natively supports a one-million-token context with dramatically lower inference cost.
Read on Outcome School ↗then come back to lock it in
Before you read, guessWhat mechanism expands the residual stream and restricts the mixing matrix to ensure a spectral norm of 1 for stable deep training?
Ten seconds, a guess, then read — a wrong guess still makes the answer stick.
What this article covers
- The Big Picture
- Two Models: DeepSeek-V4-Pro and DeepSeek-V4-Flash
- Hybrid Attention with CSA and HCA
- Manifold-Constrained Hyper-Connections (mHC)
- Muon Optimizer
- FP4 Quantization-Aware Training
- Pre-Training
- Post-Training: Specialist Training and On-Policy Distillation
- Reasoning Modes
- Putting It All Together
The article lives on outcomeschool.com. Read it there, then come back: the tutor in the margin has read it and will answer questions, and the questions below check what stayed.