4.3B seeds, no match: DeepSeek V4.1-Flash weight origins remain unproven, author says
anshulkundaje · x · 2026-11-13
Will Depue pushed back on claims he hid negative results: his report's "Limits of our investigation" section and appendix state that no signature was recovered from DeepSeek-V4.1-Flash.
V4.1-Flash shipped in September, after his results page covered the V4 models. He ran 4.3 billion seeds against every GPU layout from 16 to 512 processors with stock PyTorch, plus PyTorch's CPU generator — the same exhaustive search that found signatures for V3 and V4 — and nothing matched.
His conclusion: DeepSeek changed how it creates starting weights between V4 (April) and V4.1 (September). That alone says nothing about the chip — it could be the same NVIDIA hardware with a different generator/seeding scheme, or different hardware. Testing Huawei Ascend generators is next.
More from Models
- Dev trains 46M-param native Luganda model Pearl Chat, still collapses after 2 turns — weskambale · 2026-11-13
- RWE-Bench: Fable 5.2 leads engineering tasks at 62%, over double Opus 5.5 — Jsevillamol · 2026-11-13
- Unsloth ships GGUF quant of Inkling-2-Small for local inference — gnukeith · 2026-11-13
- User says Claude banned his account over nationality, verification appeal failed — Haunting-Hand1007 · 2026-11-13
- DeepSeek opens its app toolkit via API: web search, voice, file input and safety checks — AGI Hunt · 2026-11-13
- Open-source Claude Code skill: Opus plans, cheap models build, up to 49% cheaper — Apprehensive_Sky5123 · 2026-11-13