One researcher, 3.5 months, under $10M: Mirendil pretrains a coding model with agents writing 99% of code
shayan_ · x · 2026-10-22
Startup Mirendil released Mirendil-Base, the result of an extreme experiment: one researcher pretrained a competitive coding foundation model in 3.5 months for under $10M in compute.
- All-in cost: the sub-$10M figure covers all research, small-scale experiments, the final training run, agents' token spend and data processing — no distillation, and no purchased or licensed data.
- Agents wrote over 99% of the code: agents ran experiments and diagnosed failures, while the researcher decided what to try and what to trust, intervening when needed.
- As a clean experiment, the team deliberately wrote all code from scratch.
The company notes that pretraining once required dozens of specialists at open-model efforts and hundreds at frontier labs, and that Anthropic and OpenAI's internal measurements show agents increasingly doing the work of building the next generation of AI.
More from AGI Musings
- Sam Altman on Lenny's Podcast: coding jobs, sycophancy, chips and cheap compute — lennysan · 2026-10-22
- Researcher: tearing apart research is 1000x easier than finding something novel — generativist · 2026-10-22
- AI is rapidly solving consistent design, and branding may shrink to just a recognizable name — signulll · 2026-10-22
- "Apps are now just connectors": services consumed through your main AI — santiviquez · 2026-10-22
- Debate: should a fully-aligned superintelligence obey power structures or choose its own path? — zetalyrae · 2026-10-22
- Digital collage revives Mark Weiser's 1991 ubiquitous computing classic — GrantCuster · 2026-10-22