Kardashev-0.7: A Swarm of 32 Models? What Banbury Road Claims (and What's Missing)
Read the Portuguese translation →
Author: Prompt Engineer 48 · Watch on YouTube
Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.
Original language: English
Original video: Kardashev-0.7: A Swarm of 32 Models? What Banbury Road Claims (and What’s Missing) · Channel: Prompt Engineer 48 · 03:09 · Watch on YouTube
Topics: AI agents, AI evaluation and benchmarks, Model inference and optimization, AI models, Local and open-source models
Summary
Section titled “Summary”The video reviews Banbury Road’s claims about Kardashev 0.7, a 32-model swarm trained using a method called RL for population scaling. It describes results from an eight-model study, including a higher coverage score than individually trained models, while clarifying that the score selects the best answer from eight responses. Banbury Road’s announcement also reports a strong SuperGLUE result, very low inference costs, and a mathematical achievement, but the presenter notes these claims lack independent verification. No open weights, public API, or license were found, so the presenter calls the project promising but still at an early stage.
Key points
Section titled “Key points”- Banbury Road describes Kardashev 0.7 as a swarm of 32 distinct models trained together using RL for population scaling (RLPS).
- According to the research post, the models share a pre-trained backbone and receive their own parameter updates.
- The company says doubling the population adds about 0.6% to model-weight storage; its announcement also cites a memory requirement of about 0.03x.
- In an eight-model study, the population-trained models scored 81.70% versus 71.04% for the same models trained individually.
- The 81.70% score uses the best answer among eight responses, so it measures coverage rather than one model’s standalone performance.
- The 32-model Kardashev result appears in the announcement video; the research-post evaluation discussed is an eight-model study.
- Banbury Road reports a 49.8% SuperGLUE score for Kardashev 0.7, ahead of the three systems shown in its chart.
- The company’s chart lists Kardashev’s output-token price at 18 cents per million tokens, far below the listed comparison systems.
- The presenter has not seen proof for the claim that the system solved a 70-year-old open problem in algebraic geometry for under $500.
- The presenter could not find open weights, a public API, or a license, and says the benchmarks are from Banbury Road.
Chapters
Section titled “Chapters”| Start | Chapter | Summary |
|---|---|---|
| 00:00 | Kardashev 0.7 and the model swarm | The presenter introduces Banbury Road’s claim of a 32-model trained swarm and explains the RLPS approach. |
| 00:30 | How the models are trained and evaluated | The video describes shared backbones, storage claims, and the eight-model study’s coverage results. |
| 01:34 | Benchmark results and caveats | The presenter compares the reported SuperGLUE scores and emphasizes that the result is company-reported and limited to one benchmark. |
| 02:05 | Reported inference costs | Banbury Road’s pricing chart suggests much lower output-token costs, which the presenter says could matter if confirmed. |
| 02:37 | Unverified claims and what is missing | The presenter flags the algebraic-geometry claim as unverified and notes the absence of public weights, an API, and a license. |
Quotes
Section titled “Quotes”everything I’m going to show you comes from their own announcement. — 00:00
it is a coverage number and not what one model gets on its own. — 01:03
I would want to see independent numbers before saying more. — 02:05
I have not seen the proof, so I’m not going to tell you it’s confirmed. — 02:37
exciting idea, early days. — 03:09
Concepts and entities
Section titled “Concepts and entities”Kardashev 0.7, Banbury Road, RL for population scaling (RLPS), SuperGLUE, GPT-5.5, GLUE-3.7 max, DeepSeek V4 Pro, Qwen3.7 Max
Full transcript (original language)
Section titled “Full transcript (original language)”00:00 All right. In this video, we will look at Kardashev 0.7, a new model from a startup called Banbury Road, and they’re calling it the world’s first trained swarm made of 32 distinct models. Now, mind you, everything I’m going to show you comes from their own announcement. So, these are their claims and not my tests, but the numbers are really interesting. So, let’s go ahead and look at them. So, instead of training one giant model, they trained 32 models together, and the method is called RL for population scaling or RLPS. The idea is that these models organically develop
00:30 their own specialization, so each one gets good at different things, and they complement each other. Now, if you look here on the research post, it says, “Every member shares a pre-trained backbone and then gets its own parameter updates.” So, you’re not storing 32 full models. They say that doubling the population adds only about 0.6% to the model weight storage. And in the announcement, they say the whole thing needs about 0.03x of the required memory. Now, this part is from the research post, and it’s an eight-model result they tested on SuperGLUE, Tweet Eval, Sentiment, and 288 held-out tasks
01:03 from Musique with eight models trained as a population. The evaluation score is 81.70%, and the same eight models trained individually get 71.04%. So, that’s about 10 points from training them together. Now, mind you, this score takes the best answer out of eight responses checked against the known answers, so it is a coverage number and not what one model gets on its own. And their headline line is, “Two times the number of models gives you about the same gain as 2.02 times the parameters in a single model.” So, you can see the idea here. In the past, we scaled by
01:34 model size, and they want to scale by model count. Mind you, this is the eight-model study. The 32-model Kardashev result is only in the announcement video. So, here is the benchmark they show, SuperGLUE, which is graduate-level knowledge. Kardashev 0.7 is at 49.8%, GPT-5.5 is at 41.0, GLUE-3.7 max is at 37.9, and DeepSeek V4 Pro is at 32.6. So, on this one chart, it is ahead of all three and that’s really good, but it is one benchmark
02:05 reported by the people who built the model. So, this is a good start and I would want to see independent numbers before saying more. Now, the part that got me most is the cost. On their chart, the price per million output tokens is 18 cents for Kardashev. GPT-5.5 is $30, Qwen-3.7 Max is $4.95 and DeepSeek V4 Pro is $3.96. And you can see the scale here is logarithmic, so the gap is bigger than it looks. In the tweet, they put it as somewhere between 0.007x and 0.02x of the inference cost. So, if
02:37 this holds up, that is a really big deal for anyone running things at scale. Now, they also say they solved a 70-year-old open problem in algebraic geometry for under $500. That is a really big claim and I have not seen the proof, so I’m not going to tell you it’s confirmed. If this is true, it is amazing and I’ll be watching to see if mathematicians check it. So, what do we not know yet? I could not find open weights, a public API, or a license. The video ends with be among the first to try it. So, it looks like a waitlist for now and all the benchmarks
03:09 are from Banbury Road. So, I’d say exciting idea, early days. So, in this video, we have seen Kardashev 0.7, a swarm of 32 models that scales by model count instead of model size. If you’re interested, you can go ahead and read the research post link in the description and let me know what you think. Announcement video and charts are credited to Banbury Road. I’ll see you in another one. Check out the next video.