Ricursive's Ripple-65B hits 80.52 on TerminalBench 2.1 for $32M

Ricursive Intelligence, a lab that describes itself as building self-improving systems by co-optimizing hardware, software, and model, released Ripple-65B Preview on Wednesday, a general-purpose variant of the proprietary model behind its AI chip design platform. The headline number is 80.52 on TerminalBench 2.1, which Ricursive calls U.S. open-weight state of the art on agentic coding. The model is pretrained on 12 trillion tokens with support for up to 1M context length, and it ships on Hugging Face, Together AI, GMI Cloud, and OpenRouter, with Fireworks going live later in the day.

The benchmark table is what will get shared. Ripple-65B edges Reflection Beam 501B at 80.1, Meta Muse Spark Open 816B at 76.4, Thinking Machines Inkling 975B at 63.8, and NVIDIA Nemotron 3 Ultra 550B at 56.4. Read those parameter counts again. A 65B dense model is claiming a win over an 816B and a 975B. That is the kind of result that, if it holds up under independent eval, resets assumptions about what scale buys you on agentic coding specifically.

The cost claim is the sharper edge

Benchmark bragging is cheap. The $32M figure is not. Ricursive says it developed and trained Ripple-65B Preview for $32M in full R&D cost, “starting from scratch, including pretraining ladder development, data mixing, evals, debugging, synthetic data generation, posttraining recipe development, and the final hero pretraining and posttraining run.” The post sets that against Sam Altman’s 2023 remark that GPT-4 training exceeded $100 million and Dario Amodei’s projection that pretraining alone could reach $10-100B by 2026 to 2027. Ricursive claims its number is one-tenth the cost of the previous open-weight SOTA on agentic coding, estimated “to the best of our knowledge.”

Treat that comparison with care. It is a self-reported figure, the methodology is not disclosed, and “previous open-weight SOTA” is a moving target. The footnote in the post points at Reflection, which was reported to commit $150M per month for compute at SpaceX Colossus to train its models, with Beam taking 4 weeks to pretrain and 4 weeks to posttrain. If even roughly accurate, the gap between a $32M all-in run and a nine-figure compute commitment is the story underneath the benchmark.

How they got there: fewer rollouts, not more

The posttraining claim is the one ML engineers will actually scrutinize. Ricursive says it reached a TerminalBench 2.1 score of 80.5 with only 1.4M rollouts, “roughly 70x fewer rollouts than Reflection Beam.” It did not buy posttraining data, instead building synthetic data pipelines, quality filtering, and coverage metrics in-house. The RL recipe uses truncated importance sampling, score centering, and a linear length penalty over long-horizon synthetic environments, with task difficulty ratcheted up as the model improved. Ricursive calls the result a “steep RL climb” from a strong SFT checkpoint.

If the 70x rollout reduction is real and reproducible, it matters more than the benchmark delta. Rollout count is the cost driver in agentic RL. Labs that can climb steeply on fewer samples train cheaper and iterate faster. The open question is whether the synthetic environments encode enough genuine task diversity to transfer beyond TerminalBench, or whether the model is well-tuned to the eval it is being measured on.

Rollout count is the cost driver in agentic RL. A 70x reduction, if it holds, is worth more than the benchmark delta.

A dense model, on purpose

Ricursive chose a dense Transformer over Mixture-of-Experts, a deliberate break from most frontier labs. The post says the architecture was designed to train and serve efficiently on NVIDIA hardware, keeping infrastructure simple while holding 40% MFU through pretraining and posttraining “even with simple sharding, no custom kernels, and no quantization.” The model uses a thin-and-deep aspect ratio with a narrow d_model of 5120 and 80 total layers, overprovisioned attention heads, a higher SWA-to-global layer ratio with a smaller SWA window, plus FFN activation clipping, gated attention, global NoPE, and value residuals.

The dense choice connects to the serving story. Ricursive says the small parameter envelope lets the model run on local devices like NVIDIA DGX Spark, which have high compute FLOPs relative to memory bandwidth. A 65B dense model that serves well on a desktop-class box is a different product category from a 500B MoE that needs a datacenter.

The ladder is the real IP

The most interesting section is the model laddering log, which reads like a lab notebook. Ricursive started from the Olmo 3 ladder from the Allen Institute for AI, then iterated: v1.0 added StepFun batch size and learning rate scaling laws, v1.1 a smoother architectural scaling law, v1.2 a pre+post norm switch that was neutral at small scale but helped from 4B, v1.3 sigmoid gated attention, v1.4 a 2-stream mHC that was rejected because it did not improve at 4B, v1.5 global NoPE plus a final LR cooldown to zero, v1.6 value residuals. The final version swapped in a custom pretraining dataset with code substantially upsampled.

Ricursive says the compounded changes improved compute efficiency on code, measured by dolma_stack cross-entropy, by up to 16x at 4B scale. In plain terms, its final ladder reaches similar code performance to the Olmo 3 ladder with 16x less compute, and the post notes the baseline slope is shallower, so the gain should grow with scale. That is a credible, checkable claim, and it is the kind of finding that compounds across a lab’s future models.

What to watch

Three things. First, independent reproduction of the TerminalBench 2.1 number, since self-reported SOTA on a single agentic benchmark is the easiest claim to overfit. Second, whether the 16x compute-efficiency result holds at larger scales, because that is what determines whether Ricursive can climb toward the frontier on its budget. Third, the chip design angle: Ricursive built this model to power its own silicon work, and a lab that trains models to design the hardware that trains its models is running a loop worth watching. Ripple-65B Preview is a preview, and the weights are out. The evals will tell the rest.