Kardashev 0.7, developed by Banbury Road, is an innovative AI “train swarm” of 32 specialized models trained simultaneously using Reinforcement Learning for Population Scaling, promising improved performance and significantly lower costs compared to traditional large models. While early results show impressive benchmarks and cost efficiency, all claims currently lack independent verification, with limited public access and details pending further evaluation.
The video discusses Kardashev 0.7, a new AI model developed by the startup Banbury Road, which is described as the world’s first “train swarm” consisting of 32 distinct models. Unlike traditional approaches that focus on training one large model, Banbury Road’s method, called Reinforcement Learning for Population Scaling (RLPS), trains multiple models simultaneously. Each model specializes in different tasks, complementing one another, and they share a common backbone with individual parameter updates. This approach reportedly requires significantly less memory, with only a 0.6% increase in storage when doubling the population of models.
An earlier study with eight models showed promising results, where the population-trained models achieved an evaluation score of 81.7% on various tasks, compared to 71.04% when trained individually. This improvement is attributed to the models’ complementary specializations and the ability to select the best answer from multiple outputs. The key insight from Banbury Road is that increasing the number of models can yield similar performance gains as increasing the size of a single model, marking a shift in scaling strategy from model size to model count.
The video highlights the performance of Kardashev 0.7 on the Super GPQA benchmark, which tests graduate-level knowledge. According to Banbury Road’s reported results, Kardashev 0.7 outperforms other models like GPT 5.5, Qen 3.7 Max, and Deepseek V4 Pro, achieving a score of 49.8%. However, these results come solely from the developers, and independent verification is still pending. The presenter emphasizes the need for external benchmarks to confirm these claims.
One of the most striking claims is the cost efficiency of Kardashev 0.7. The price per million output tokens is reported at just 18 cents, which is dramatically lower than competitors like GPT 5.5 at $30 and others around $4 to $5. This cost advantage, if validated, could be transformative for large-scale AI deployments. Additionally, Banbury Road claims to have solved a longstanding problem in algebraic geometry for under $500, though this extraordinary claim remains unverified and awaits scrutiny from the mathematical community.
Despite the exciting potential, several important details remain unknown. There is no public release of model weights, API access, or licensing information, and access to the model currently appears to be via a waitlist. All benchmarks and claims come from Banbury Road itself, so independent testing is necessary to fully assess the model’s capabilities. Overall, Kardashev 0.7 represents an innovative approach to AI scaling with promising early results, but it is still in the early stages of development and evaluation.