Efficiency Claims in Large-Scale Model Training

The startup Mirendil has released details regarding its recent large-scale pretraining efforts, claiming to have developed a model comparable in performance to Kimi K3. According to reports circulating on social media platforms such as Reddit, the startup achieved this milestone with a single developer and a reported compute expenditure of $10 million.

The company’s own documentation, hosted on its website, suggests a shift in the traditional paradigm of large-scale pretraining, implying that such results can be achieved without the massive, dedicated pretraining teams typically required by major industry players.

Discrepancies in Cost Accounting

The reported $10 million figure has drawn significant scrutiny from industry observers regarding the methodology used to calculate training costs. Discussions on Reddit highlight a lack of consensus on what constitutes “compute cost.” Some commentators suggest the figure may refer strictly to electricity and operational expenses, while others argue that excluding hardware costs provides an incomplete picture of the total capital required for such an undertaking.

While Mirendil claims a $10 million expenditure, some analysts have noted that this figure may not represent a significant outlier in the current market. For instance, users on Reddit pointed to Xiaomi’s reported training costs for Mimo 2.6 Pro, which were cited at under $4 million, suggesting that a sub-$10 million budget is within the realm of possibility for high-level model training.

Skepticism Regarding Scale and Resource Allocation

The claim that a single developer could manage a training run of this magnitude has met with varying degrees of skepticism. While some users questioned the feasibility of such a lean operation, others focused on the broader economic context of AI development.

Debates have emerged regarding the total capital required to reach current performance benchmarks. While some speculative comments suggested astronomical costs reaching into the trillions, other observers noted that industry estimates for training runs of recent high-performing models are more likely to be in the billion-dollar range.

At this stage, the specific technical architecture and the exact composition of the compute resources used by Mirendil remain unverified by independent third-party audits. The company’s claims represent a significant departure from the resource-heavy development cycles currently dominated by large-scale technology firms.