A 65B dense model tops US open-weight coding, built for $32M

Ricursive Intelligence, a lab that builds AI for chip design, has released Ripple-65B Preview, an open-weight model it says sets a U.S. open-weight record on agentic coding: 80.52 on TerminalBench 2.1, with a full R&D cost of $32 million.

A dense 65B model that outscores far larger rivals

According to Ricursive, Ripple-65B Preview is a general-purpose variant of the proprietary model behind its chip design platform. It was pretrained on 12 trillion tokens and supports context of up to 1M tokens. The company says it trained the model from scratch.

On TerminalBench 2.1, Ricursive reports 80.52. It lists these U.S. open-weight comparisons: 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. The margin over Beam is under half a point, so the lead is narrow, though the parameter gap is large.

The architecture is a modernized dense Transformer, not a Mixture-of-Experts design. Ricursive describes a thin-and-deep layout with a model width of 5120 and 80 layers. It also lists gated attention, global NoPE, value residuals and FFN activation clipping. The company says it kept about 40% MFU through pretraining and posttraining with simple sharding, no custom kernels and no quantization.

The post also describes a short, efficient reinforcement learning phase. Ricursive says it reached its score with 1.4M rollouts, roughly 70 times fewer than Reflection Beam. It says it bought no posttraining data and built its own synthetic data pipelines and long-horizon environments.

Where to get it and how to run it

Ricursive says the model is on Hugging Face, Together AI, GMI Cloud and OpenRouter, with Fireworks going live on the day of the announcement. The company also says the small parameter count lets it run on local hardware such as NVIDIA DGX Spark. That machine has high compute but comparatively low memory bandwidth, which suits a dense model of this size.

The page does not state a license, pricing for the hosted endpoints or the weight precision. Anyone planning commercial use will need to check the Hugging Face model card first. The “Preview” label also suggests the release is not final, though the company does not say what changes.

Ten times cheaper than the last leader, by its own estimate

The economics are the center of Ricursive’s pitch. It puts the full cost at $32 million, covering ladder experiments, data mixing, evals, debugging, synthetic data, recipe development and the final runs. It calls this one-tenth of the previous open-weight leader’s cost, and labels the figure as an estimate to the best of its knowledge.

That comparison rests on a footnote. Ricursive says Reflection was reported to have committed $150 million a month for compute at SpaceX Colossus, and that Beam took four weeks to pretrain and four weeks to posttrain. Eight weeks at that rate would be roughly $300 million, which is the arithmetic behind the tenfold claim. A monthly commitment is not necessarily what was spent, and the source gives no confirmation from Reflection.

Two outside details help frame the comparison. Technical coverage by DevX describes Beam as a 501B Mixture-of-Experts model that activates about 23B parameters per token. So the per-token compute gap with a dense 65B model is smaller than the headline parameter counts suggest, and dense models pay for every parameter on every token. Separately, a Brazilian report on Beam’s launch, Hora de Codar, says Artificial Analysis gave Beam an Intelligence Index of 30, second among U.S. open-weight models behind Muse Spark Open at 39. That index is a broad measure, not a coding-only test, and it shows that a TerminalBench lead does not settle overall ranking. I could not verify that figure independently.

The same sources disagree on whether Beam’s weights were fully released, so cost and openness comparisons with it should be treated cautiously.

What the announcement does not establish

Every benchmark number comes from Ricursive’s own post. The page does not link independent reproductions, give the evaluation harness settings or say how rivals’ scores were obtained. The $32 million figure is also self-reported, and the post does not itemize it.

The company frames the release as a step toward a loop in which models help design the chips that train them. An internal variant already powers its chip design platform, and it says it plans to scale toward the frontier. It also credits earlier work from the Allen Institute for AI’s Olmo team, Prime Intellect, RadixArk and NVIDIA. For developers, the practical test is simple: the weights are public, so the TerminalBench claim can now be checked by others.