Mirendil Releases Mirendil-Base Code Foundation Model, Pretraining Completed in 3.5 Months
Mirendil has released its code foundation model Mirendil-Base, pretrained from scratch by one researcher leading a group of Agents over 3.5 months at a cost of under $10 million, producing a model with 100–500 billion parameters. It used only public data—no distillation, no synthetic data, and no purchased data—and more than 99% of the code was written by Agents. On code prediction loss it outperforms GLM-5.3, DeepSeek-V4-Pro-Base, and Nemotron-550B, but falls short of Kimi-K3. The model does not release its weights, has no API, and has no task evaluations such as SWE-bench yet; the author views this as more of an efficiency win for a senior researcher plus Agents, and the method relies on third-party Coding Agents and is easy to replicate.
Why read this
AI score 67
This piece lays out the cost, data sources, and comparison results of a small team's pretraining experiment, offering a basis for judging the practical boundaries of Agent participation in foundation model training.
Topics
Body · AI translation
One researcher + a group of Agents, 3.5 months, less than $10 million, pretraining a 100–500 billion parameter code foundation model from scratch. Mirendil released Mirendil-Base today. It started from an empty code repository, using only public data—no distillation, no synthetic data, no purchased data—and over 99% of the code was written by Agents. On code prediction loss, it outperforms GLM-5.3, DeepSeek-V4-Pro-Base, and Nemotron-550B, but still falls short of Kimi-K3 👀
Of course, it does not release weights, has no API, and there are not yet task evaluations like SWE-bench. But the signal is already clear: the bottleneck in pretraining is shifting from "people" to compute, data, and the judgment of the person directing the Agents. This means three things:
1. Chip design teams, robotics labs, and even hospitals may all be able to train their own dedicated foundation models on their own data, rather than merely renting general-purpose models for fine-tuning;
2. Organizational expansion no longer relies only on hiring people—adding compute can amplify each person's output;
3. A researcher's daily work will look more like running a small lab: setting direction and allocating scarce compute.
But my judgment is: this is an efficiency victory for senior researchers + Agents, not evidence that the Mirendil model itself is that strong. This approach relies heavily on third-party Coding Agents, and the moat is only the researcher's judgment and compute contracts—it is very easy to replicate. And according to reports, Mirendil is raising funding at a $5 billion valuation... No wonder some joke that this looks more like an ad for Codex or Claude 😅