DeepSeek V4.1 Leak: 552B New Architecture, HBM Needs Cut 3.9x, Huge Benchmark Gains

teortaxesTex · x · 2026-09-10

Unofficial benchmarks for DeepSeek V4.1 show a new 552B-parameter Causal-Encoder-Decoder architecture (8B active / 16B output) posting strong numbers: 31.2 on TerminalBench 4.0, 88.1 on CyberGym, 15.3 on ExploitGym, 63.9 on HLE with tools, and 54.8 on Automation-Bench.

On efficiency, HBM requirements are reportedly slashed 3.9x and SSD usage 8x versus V4-Flash, with KV cache efficiency still unmatched six months after V4's arch debuted. The poster estimates >90% inference margins and quips that an American lab with these results would be labeled "recursive self-improvement." Figures are unverified leaks.

Related event: DeepSeek V4.1 and V4.1 Flash Benchmarks and Architecture Allegedly Leak(8 posts)→

Original post →

More from Models

Models channel →