Ricursive's Ripple-65B claims US open-weight SOTA on agentic coding with 70x fewer rollouts

Ricursive Intelligence, a startup whose stated mission pairs AI for chip design with chip design for AI, has published a preview of Ripple-65B, a model it calls the strongest US open-weight entry on agentic coding benchmarks. The 65B in the name points to a 65-billion-parameter model. The announcement reached the Hacker News front page on November 4.

The headline numbers

According to Ricursive, Ripple-65B Preview scores 80.5 on TerminalBench 2.1, a benchmark that measures how well AI agents complete real work in a terminal environment. The company says it reached that score with roughly 1.4 million RL rollouts, which it estimates is about 70 times fewer than Reflection Beam, the model it identifies as the previous state of the art on the benchmark.

The claim is twofold: a higher score than the prior leader, achieved with a small fraction of the reinforcement learning compute. Ricursive frames that gap as evidence that the quality of its training recipe, rather than raw scale, decided the outcome.

A recipe built for compute efficiency

Ricursive says it deliberately avoided an expensive, brute-force RL climb and instead iterated on both its supervised fine-tuning (SFT) and reinforcement learning setups with compute efficiency as the explicit goal. The RL phase started from a strong SFT checkpoint and, in the company's description, produced a steep improvement curve from there.

Two details stand out. First, Ricursive says it purchased no posttraining data at all, instead building its own synthetic data pipelines, with quality filtering and coverage metrics to keep track of what the training distribution actually contained. Second, its mixed RL recipe combines truncated importance sampling, score centering and a linear length penalty, applied in what the company describes as high-quality, long-horizon synthetic environments.

The task mix was also dynamic. Ricursive says RL tasks were tuned to the model's current capabilities, with harder tasks added continuously as the model improved — effectively a curriculum that grew alongside the model rather than a fixed task set.

Claims worth treating as claims

The figures come from Ricursive's own announcement rather than an independent evaluation, and the release is explicitly a preview. The 70x rollout comparison is also the company's own accounting. Rollout counts are a useful but incomplete proxy for training cost: rollout length, environment cost, model size and infrastructure all affect the total compute bill, and none of those factors are broken out in the summary currently available. Independent benchmark runs and the model's release terms will determine how much of the claim holds.

Why it matters

If the numbers survive outside evaluation, two things follow. The first is competitive: a US lab would hold the open-weight state of the art on agentic coding, a category where open models have been steadily closing the gap with closed systems. The second is economic: reinforcement learning on long-horizon tasks is among the most expensive stages of training a coding agent, and a roughly 70x reduction in rollouts would change who can afford to train such models at all.

The method may matter as much as the score. Ricursive's approach — generating posttraining data in house, filtering it for quality and coverage, and escalating task difficulty as the model improves — is a repeatable recipe rather than a one-off scaling result, and one that smaller teams could in principle copy.

The company's own framing is also worth noting. Ricursive describes its work as recursive self-improvement, with AI accelerating chip design and better chips feeding back into AI training. A compute-efficient training method fits that loop directly: the cheaper each improvement cycle becomes, the faster the cycle can run.