Mirendil-Base: one researcher pretrained a 100B+ coding base model in 3.5 months for under $10M
bneyshabur · x · 2026-10-22
Mirendil, announced less than 4 months ago with a mission to solve Recursive Self-Improvement, published results of a controlled experiment: pretraining at scale without a pretraining team.
- Headline numbers: Mirendil-Base was built from scratch by one researcher in 3.5 months for under $10M in compute, including all small-scale experiments and the final run. It sits in the 100B–500B parameter range; Base-Large is 500B–2T and Base-XL is 2T–10T.
- Evals (cross-entropy in bits per byte on held-out code, lower is better): Base-Large hits 0.2177 (80% trained) and Base-XL 0.2181 (19% trained), both beating Kimi-K3 at 0.2387; Base at 0.2803 outperforms Nemotron-550B-Base, GLM-5.3, and DeepSeek-V4-Pro-Base.
- Thesis: a single AI-empowered researcher can now do work that once required an entire research organization; the bottleneck shifts from team size to the ambition and judgment of the person directing the systems.
Independent evaluations, post-training benchmarks and API access are promised soon.
More from AGI Musings
- Economists push back on Pangram AI-detection policing of academic writing — paulnovosad · 2026-10-22
- MLB's ABS challenge system shows where sports draws the line on machine decisions — lmoroney · 2026-10-22
- What MLB's ABS challenge system and NBA's sensor ball teach us about handing decisions to machines — lmoroney · 2026-10-22
- Dario Amodei's stark take: AI threatens identity and the tax base, not just wages — coherence · 2026-10-22
- Wiblin's plea to AI lab staff: reassess now whether your work delays or accelerates RSI — geoffreyirving · 2026-10-22
- Competition-driven AI progress could make science uncontrollable and incomprehensible — ShahabBakht · 2026-10-22