DeepSeek near-perfect on three rounds of obscure trivia while MiniMax fabricates long-tail facts
teortaxesTex · x · 2026-10-20
- Using obscure trivia quizzes (generated by Sonnet, answers verified via web search), @teortaxesTex benchmarked DeepSeek vs MiniMax M3.1-Flash across three rounds.
- Round 1 (24 pts): DeepSeek 24/24, MiniMax 14/24 — MiniMax gets headline facts right but fabricates second-order details (heights, ages, causes of death).
- Round 2 (31 sub-answers): DeepSeek 31/31, MiniMax 15/31, inventing nonexistent hills, chemists, and paper stories, and taking far longer (e.g. 2m49s vs 8s).
- Round 3 (37 sub-answers): DeepSeek 34/37, MiniMax 23/37; MiniMax burned its full 20K-token budget without answering the Turing question.
Takeaways: DeepSeek's failure mode is narrow — flawless on digit-level facts, misses limited to single secondary fields in obscure narratives. MiniMax is strong on dense, widely replicated tables but weak on narrative long-tail, replacing whole stories with fabricated ones. The author also admitted one of his own answer keys was wrong — DeepSeek's answer was better than his key. In an earlier test, MiniMax V4.1 Flash scored 100% on perturbed easier variants vs 85% on originals, while DeepSeek's Whale aced both.
Related event: DeepSeek Aces Obscure-Facts Quiz, Crushes MiniMax in Head-to-Head Test(2 posts)→
More from Models
- Inception's Mercury diffusion LLM wows in speed test with essay on parallel thinking — gkavek · 2026-10-20
- DeepSeek open-sources DeepInfer, the production inference framework behind DeepSeek-V4.1-Flash — deepseek_ai · 2026-10-20
- DeepSeek-V4.1-Flash-1017 Hits Qoder Day One, Free at 0.0x Credits Until Oct 24 — alifcoder · 2026-10-20
- DeepSeek-V4.1-Flash quantized to 2-4 bit mixed precision for dual RTX PRO 6000 with 0.1306 KL divergence — TheZachMueller · 2026-10-20
- Vals AI: DeepSeek V4.1 Flash Beats GPT-6 Sol on Offensive Cybersecurity Benchmark — Hesamation · 2026-10-20
- MemTensor releases MemT-1.0, a memory-native architecture embedding continuous memory into model computation — MemTensor · 2026-10-20