A new paper offers a striking set of numbers: 801-30B-A3B-Base was trained over three months on 11.8 trillion tokens using a mixture-of-experts architecture with 30 billion total parameters and A3B active parameters. In our view, what matters more than a single benchmark result is that it moves model self-improvement out of post-training and into frontier pretraining, a stage that has previously been difficult to reproduce.

What this is

Recursive self-improvement, or RSI—a training process in which models participate in improving their successors—has largely occurred during post-training: teams first train a base model, then experiment through reinforcement learning or supervised fine-tuning. This paper applies the approach to training a base model from scratch and breaks an end-to-end process that cannot be fully rerun into smaller, verifiable loops that address problems exposed in the data pipeline, training infrastructure, and individual steps.

The result is the open base model 801-30B-A3B-Base and a lightweight supervised fine-tuned version trained on only six billion tokens, without reinforcement learning. The paper says the former outperformed Qwen3-30B-A3B and Nemotron-3.5-Lightning-30B-A3B in like-for-like comparisons, while the lightweight version achieved the highest average score across four agent benchmarks. The full model and data will be released progressively, so independent replication is still pending.

Industry view

We believe the meaningful signal is not the sweeping claim that “AI has begun to build better AI,” but that model development is shifting from isolated algorithmic competition to competition over complete production systems. The paper explicitly notes that a single frontier pretraining run may consume hundreds of thousands of GPU hours, last for months, and be impossible to reproduce exactly. The object of improvement is no longer a model alone, but the data, infrastructure, and collaborative processes surrounding it.

The risks are equally clear. First, whether the smaller loops genuinely improve the final training outcome still depends on the paper’s disclosure of data and code and on replication. Second, the lightweight fine-tuned model’s performance on agent tests cannot be treated as equivalent to the capabilities of a finished product. Third, 801’s comparisons use model, training, and testing conditions chosen by its authors and may not hold against closed models. More importantly, work of this kind requires sustained access to compute, training data, and engineering teams; in the short term, it will not automatically lower the barrier to frontier capabilities for every company.

Impact on regular people

For enterprise IT: As organizations procure models, data governance, training pipelines, and continuous evaluation will become even more important. Public leaderboard rankings alone will no longer be enough to judge whether a vendor can iterate reliably.

For individual careers: Models that can invoke code, retrieve information, and execute multi-step tasks are getting stronger. But these capabilities currently manifest more often as benchmark scores and open models; their reliability in real jobs must still be tested task by task.

For the consumer market: If the planned releases materialize, developers and small and medium-sized businesses may gain more open and locally deployable options to research. In the short term, the direct impact will remain concentrated in the model ecosystem rather than immediately changing consumer-device prices or features.