AI R&D startup Mirendil announced Mirendil‑Base, a pretrained coding-focused large language model developed by a single researcher in about 3.5 months using less than $10 million of compute. Mirendil says that figure covers all the research, small experiments and the final training run.
What they did
- All implementation work was done from scratch by the researcher, building on libraries such as PyTorch and JAX.
- Training data came exclusively from publicly available sources, according to Mirendil.
- Model scale tiers: Mirendil‑Base is in the 100 billion–500 billion parameter range; Mirendil‑Base‑Large in the 500 billion–2 trillion range; Mirendil‑Base‑XL in the 2 trillion–10 trillion range.
- Progress: Mirendil‑Base‑Large is reported to be about 80% through its training run; Mirendil‑Base‑XL is about 19% through and is already approaching Large’s current score.
Evaluation and comparisons
Mirendil‑Base reportedly outperforms models such as Nemotron‑550B‑Base, GLM‑5.3, and DeepSeek‑V4‑Pro‑Base, but underperforms compared with Kimi‑K3. Mirendil states that both Mirendil‑Base‑Large and Mirendil‑Base‑XL exceed Kimi‑K3 in their internal comparisons.
Why this matters
Mirendil frames the release not primarily as a novel technical breakthrough but as a proof‑of‑concept for how AI R&D optimization can change development workflows. The company argues the case that a single human, working in partnership with AI tools, can reach results that previously would have required teams of many people. They note that the capability of AI tools to assist in model development has grown dramatically over the last year, and that models that help build other models make subsequent models cheaper and faster to produce.
Mirendil is careful to add that they do not claim to have reached fully autonomous recursive self‑improvement, but believe they are closer to that regime than before.
Implications
The Mirendil‑Base announcement illustrates a broader trend: with powerful tooling and large amounts of compute, small teams or even individuals can approach the frontier of open‑weight models for particular size classes. If this pattern continues, it may shift how AI research is organized and accelerated in the coming years.
Further information
Mirendil’s writeup titled "Pretraining at scale without a pretraining team" provides more detail on their process and findings.
