ripple 65b preview: Ricursive Claims US Open-Weight Coding Lead

The ripple 65b preview is the kind of open-weight drop that makes AI Twitter do math. Ricursive Intelligence, a lab that builds AI chip-design systems, released Ripple-65B Preview as Apache 2.0 open weights on November 8, claiming a US open-weight record of 80.52 on Terminal-Bench 2.1 for agentic coding. The company says the whole R&D bill was $32 million, and that the model is small enough to run on local hardware.

What actually happened

Ricursive's announcement describes the release as a general-purpose variant of the proprietary model that powers its chip design platform. The post says Ripple was trained from scratch and optimized for the company's training and serving setup, with a dense Transformer rather than a Mixture-of-Experts design. A dense model uses all its parameters on every token, which tends to keep latency predictable, while MoE models like those we covered in inkling 2 small and ernie 5 5 flash base activate only a slice per token.

According to the company, Ripple-65B Preview scores 80.52 on Terminal-Bench 2.1 and beats several much larger US open-weight models it lists: 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. Those comparison scores come from Ricursive's own post. The Independent AI's write-up flags the claim as self-reported, and we have not seen an independent reproduction yet.

The $32 million claim

The headline number is cost. Ricursive says it built the model for $32 million of full R&D spend, which it describes as about one-tenth the cost of the previous open-weight leader on agentic coding. It adds that the figure includes pretraining experiments, data mixing, evals, debugging, synthetic data generation and post-training recipe work, not just the final run.

That comparison rests on an estimate. Ricursive's footnote says other labs were reported to commit $150 million per month for compute, and that Reflection's Beam took four weeks of pretraining and four weeks of post-training. The lab also says it reached its score with about 1.4 million reinforcement-learning rollouts, roughly 70 times fewer than Reflection Beam. Treat all cost and rollout comparisons as the vendor's framing until someone replicates them.

What is actually inside

The architecture notes in the post are unusually concrete. Ripple is "thin and deep," with a model width of 5,120 and 80 layers, and uses gated attention, global NoPE, value residuals and sliding-window attention mixed with global layers. Ricursive says it started from the Olmo 3 model ladder and credits the Allen Institute for AI, Prime Intellect, RadixArk and NVIDIA for open-source groundwork. It says pretraining ran at about 40% MFU without custom kernels or quantization.

For readers who are not training models, the practical part is size. A 65B dense model is far smaller than the 500B-plus models on the comparison list, and Ricursive says it can run on local devices such as NVIDIA DGX Spark. That does not mean it runs on a laptop; it means a well-equipped desk setup, not a data center, is in range.

Where to be skeptical

The ripple 65b preview headline is a claim, not a verdict. Terminal-Bench is one benchmark, and "US open-weight state of the art" is a narrow label: it excludes non-US models and closed ones. Scores on agentic coding evals can shift with the harness, the prompt format and the number of attempts. The word "Preview" in the name also signals that the company expects changes.

Another thing to watch is the license and the claim that the model also powers real chip-design work. The released variant is a general-purpose version, not the proprietary one. So the public checkpoint tells you how Ripple codes, not how well Ricursive's internal system designs chips.

What to do as a reader

If you build with open models, try the Ripple-65B Preview page on Ricursive's site and the Hugging Face weights on your own tasks before trusting any leaderboard, including the one in the announcement. Run a small test on your repository, compare cost and latency with what you use today, and read the license.

If you only follow AI news, the safe takeaway on the ripple 65b preview is modest: a small lab claims a strong coding result at low cost, and the claim is waiting for outside tests. Nothing here is investment advice, and benchmark scores are not guarantees of real-world performance.