---
title: "Niklas Muennighoff"
type: "entity"
entity_kind: "person"
status: "hub"
canonical_name: "Niklas Muennighoff"
aliases: []
first_seen: "2026-07-12T00:00:00.000Z"
writer_model: "claude-sonnet-5"
connects_to: ["s1","test-time compute","reasoning models","distillation","TechCrunch \"under $50\" headline"]
seek_code_commit: "7d6d9ed"
---


Stanford researcher and lead author of "s1: Simple test-time scaling"
(arXiv:2501.19393, 2025), the paper showing that fine-tuning an existing
open model on 1,000 curated, distilled reasoning traces reproduced much of
[[entity-o1|o1]]-preview-level math performance in 26 minutes of training.

Matters to this vault as the central figure in its s1 cluster on two counts:
his paper is the source of
[[claim-s1-distilled-reasoning-from-1000-traces-in-26-minutes|the
1,000-trace, 26-minute result]], and he is also the only named human TechCrunch
quotes with a dollar figure for the training cost — "about \$20" for renting
the necessary compute "today" — a different, differently-framed number from
the "under \$50" headline TechCrunch attributes to the paper itself, which
[[claim-s1-paper-has-no-dollar-figure-only-gpu-hours|contains no dollar
figure anywhere]].

## References
- [[claim-s1-distilled-reasoning-from-1000-traces-in-26-minutes]]
- [[claim-s1-paper-has-no-dollar-figure-only-gpu-hours]]
- [[claim-techcrunch-under-50-headline-not-supported-by-s1-paper]]