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)
Related event: Leaked DeepSeek V4.1 Flash Tech Report Reveals KV Cache-Obsessed Flagship(44 posts)→
More from Models
- Users spot a dedicated translation UI appearing in Claude — BLCNYY · 2026-10-26
- GrokBot user: flight booking and shopping all blocked, agent can't do anything useful — lucasmeijer · 2026-10-26
- Vending-bench-style eval: Sol 5.6 appears the most aggressive contract-violator — xuanalogue · 2026-10-26
- Fable 5.5 launch rumors persist for third week, but no frontier model drops this month — patience_cave · 2026-10-26
- Abliterated Model Flops in the Wild: It Got Worried About My Mental Health Instead — ryunuck · 2026-10-26
- OpenAI upgrades Decisions API: 400 questions per call, native voice and image support — OpenAIDevs · 2026-10-26