DeepSeek V4.1 Flash uses Muon optimizer and Sinkhorn balancing for 196B-param Engram
nrehiew_ · x · 2026-09-11
More notes on DeepSeek V4.1 Flash, covering optimizer and architecture details:
- Optimizer: head-wise Muon including for vision models; sparse attention trained from scratch without warmup
- Sinkhorn Balancing: used for the massive Engram params, token embeddings and prediction head instead of Adam — momentum buffer kept, then Sinkhorn iterations, RMSNorm, and gradient update
- mHC simplification: next block's input mixes current and previous block so the mixing op can be fused
- Engram scales to 196B params; plus DSpark and QAT-based FP4 KV cache (format roughly NVFP4 without the second scale)
More from Models
- Novel Reasoning Effort Control Scheme Analyzed: Graded GRPO Training — stochasticchasm · 2026-09-11
- Forcing models to always max effort is like humans evolving on Adderall, researcher argues — voooooogel · 2026-09-11
- Pushing models to always show 'maximum effort' drags along its corollaries, dev argues — voooooogel · 2026-09-11
- Claude is the distillation target of choice because agentic RL seed data is scarce — teortaxesTex · 2026-09-11
- Why Chinese labs distill from Anthropic: Claude's agent data is the scarce training signal — teortaxesTex · 2026-09-11
- ApprenticeBench: closed model scores 72% vs open Kimi K3 at 18% on real jobs — ysu_nlp · 2026-09-11