Banbury Road AI has announced Kardashev-0.7, which it describes as a trained swarm of 32 distinct models. In a post on X, @MLCatttt says the models were trained with RL for Population Scaling (RLPS) to develop complementary specializations, and reports inference costs of 0.007x to 0.02x and memory use of 0.03x relative to an unspecified reference system.

The announcement also invites readers to join an API beta waitlist. It does not specify signup requirements, access timing, pricing or general availability.

Video: Opening post video presents Kardashev-0.7 as a 32-model trained swarm.

Load the post to view it as published on X. X may receive connection data.

View the original post on X ↗

What Kardashev-0.7 is

Reinforcement learning (RL) is a training method in which a system improves its behavior using feedback. Banbury Road AI names its approach RL for Population Scaling, or RLPS, but does not define the objective or describe the training procedure in the announcement.

The team describes 32 separate models developing different strengths that complement one another. That differs from simply deploying multiple copies of one model. It also differs from specialization inside a single model: here, the specialization is described as emerging across a population of distinct models. The announcement does not say how the swarm routes tasks, coordinates the models or combines their outputs during inference.

The source also distinguishes Kardashev-0.7 from a 16-model system discussed in the company’s blog. Another reply says a paper shared in the thread studies a swarm that searches convolutional neural network architectures, while Kardashev-0.7 is described as learning complementary capabilities across a population of large language models.

The reported cost and memory figures

Banbury Road AI reports that Kardashev-0.7 delivers what the announcement calls “frontier performance” at 0.007x to 0.02x of the inference cost and 0.03x of the required memory. These are relative ratios, not dollar costs or a stated amount of memory. The announcement does not identify the baseline model, workload, hardware, model sizes, latency or measurement method.

The announcement does not provide benchmark scores, test datasets, quality metrics or a comparative evaluation. Without those details, readers cannot determine whether the reported efficiency holds for particular tasks or how it relates to output quality. The figures should therefore be treated as reported results rather than independently established performance.

How to join the API beta waitlist

The announcement directs interested users to the Kardashev API beta waitlist. It says people can join to be among the first to try Kardashev-0.7, but it does not establish when access will begin, who will be admitted or whether the API is currently usable.

What remains unanswered

The announcement leaves several practical and technical questions open:

  • How are the 32 models trained, coordinated and selected for a task?

  • What does the RLPS objective optimize, and how does specialization emerge?

  • What baseline supports the inference-cost and memory ratios?

  • What benchmarks, datasets and hardware produced those measurements?

  • What are the models’ sizes, capabilities, latency and API limits?

  • Will the beta include documentation, pricing or production access?

Kardashev-0.7 is an announced 32-model trained swarm with an API beta waitlist. Its reported efficiency figures and specialization approach still need detailed technical documentation and independent evaluation before readers can assess how it performs in practice.

Sources