Rumor: DeepSeek V4.1 Pro not coming soon, sparking >1T-scale architecture doubts
teortaxesTex · x · 2026-10-23
An unverified leak says DeepSeek won't ship a V4.1 Pro in the short term. Commenters speculate a flagship training setback and question whether this broad architecture family works at >1T scale, noting V4.1-Flash is strong — a risk for compute-poor, talent-rich labs that can't afford a scaling ladder. Attached paper figure shows V4.1-Flash's 40-layer design: 20-layer causal encoder + 20-layer decoder, all feed-forward layers using DeepSeekMoE, first two encoder layers with sliding-window attention and the rest Compressed Sparse Attention 2 (CSA2) with configurable ratio/mode, plus Single-Pass mHC, Engram, DSpark, and a Hierarchical Sparse Indexer; the paper positions Flash as a new starting point for joint scaling of architecture, pre-training and post-training.
More from Models
- TypeSafe's Jev: a "Decision Model" LLM That Returns Floats at $0.042/M Input Tokens — ariG23498 · 2026-10-23
- Microsoft's MAI-Base-2 scaling ladder pegs the L42 run at ~1.8×10^23 FLOPs — teortaxesTex · 2026-10-23
- Local model agentic coding benchmark: 27B matches 125B, tool-calling is the lever — Joralk_ · 2026-10-23
- NVIDIA open-sources PersonaPlex-7B, a full-duplex voice AI model that listens and talks simultaneously — solyarisoftware · 2026-10-23
- NVIDIA's KumoRFM-2: a foundation model that predicts churn, fraud and demand straight from relational databases — solyarisoftware · 2026-10-23
- Gemini Analyzes Whether a Simple 'Hi' Is a Jailbreak Attempt in Its Thinking Steps — OldIntroduction2909 · 2026-10-23