← All model news

Ripple-65B Preview releases 65B dense weights under Apache 2.0, self-reports 80.52 on TerminalBench 2.1

The Decoder reported on November 6, 2026 that Ricursive released Ripple-65B Preview: a 65-billion-parameter dense model that the company says scores 80.52 on TerminalBench 2.1, the best result for a US open model, with weights on Hugging Face under Apache 2.0; several other benchmarks trail Inkling 2 Small.

Published Updated
Cover from the original report — Ripple-65B Preview releases 65B dense weights under Apache 2.0, self-reports 80.52 on TerminalBench 2.1
Cover from the original reportOriginal report ↗

What changed

The Decoder reported on November 6, 2026 that AI chip design company Ricursive Intelligence trained and released the open language model Ripple-65B Preview, pretrained on 12 trillion tokens. According to the report, the company self-reports a score of 80.52 on the TerminalBench 2.1 coding benchmark and calls it the best result for a US open model on that benchmark; the score was obtained partly using Terminus 2 and partly using OpenCode as the harness. The report also lists three lower self-reported results: SWE-Bench Pro v2.0 Hard 79.1%, DeepSWE v1.1 44.3%, and MCP Atlas 71.67%, all below the 86.3%, 72.1% and 85.4/85.0 posted by Inkling 2 Small released around the same time. The model is a 65-billion-parameter dense Transformer that does not take the more common mixture-of-experts route, and is said to be optimized for Nvidia hardware. According to the company, total R&D and training spending was about 32 million US dollars. Weights are available on Hugging Face under the Apache 2.0 license. The report does not mention third-party reproduction.

What it means for developers

For developers planning to integrate or self-host it, what often determines whether this model is usable is not the number 80.52 but how it was measured. The report states explicitly that the score was obtained partly with Terminus 2 and partly with OpenCode as the harness, which is vendor-reported methodology: the score corresponds to the model combined with a specific agent framework, and extrapolating it directly to your own tool-calling loop is an inference, not a conclusion of the report. The second difference is the architectural route: the report says only that this is a 65-billion-parameter dense model that does not use MoE, which implies full activation per token, so the ratio of activated to total parameters differs clearly from an MoE of the same size. At roughly two bytes per parameter for BF16, full weights would be on the order of 130GB—an arithmetic inference from the published parameter count, not source data—so a single 80GB-class GPU cannot hold the full BF16 weights, and multiple GPUs or quantization would usually be needed; but the report gives no context length, quantized versions, minimum VRAM, or inference framework support, and those are the main variables in self-hosting cost that cannot be derived from parameter count alone. The third is the license boundary: the report says only that the weights are provided under Apache 2.0, and does not say whether training code, training data and accompanying tools are also open, so extending a "commercially usable" conclusion from the weights to the whole project is not safe. In addition, the report offers only one comparison point, Inkling 2 Small, and Ripple trails on three of those measures, so a single TerminalBench 2.1 leaderboard is not enough to support an overall capability judgment.

Integration and evaluation

The following is conditional analysis and advice, not reported fact. First, given the reported change that the TerminalBench 2.1 score was obtained with two harnesses, Terminus 2 and OpenCode, the choice is whether to first put it in a candidate pool for shadow comparison or to directly replace your current coding model; the premise for changing that choice is how closely your agent framework resembles those two harnesses—the larger the difference, the harder the score is to transfer. A testable standard: run a comparison with your own harness on the same batch of internal tickets, recording end-to-end pass rate, average number of tool-call turns and the distribution of failure types; only when the pass rate is no worse than your current model and the turn count has not risen noticeably is there reason to consider switching. Second, given the reported facts that it is 65B dense, optimized for Nvidia, and lacks deployment details, the choice is whether to deploy directly on your own hardware or first validate results through an external service before deciding on purchases; the premise is your available VRAM per machine and target context length. The standard: measure first-token latency and stable throughput at the target context length, and compare pass rates before and after quantization on the same batch of tasks, requiring the drop to stay within a threshold you set in advance; otherwise you need more GPUs, or a fallback plan for long-context scenarios. Third, given the reported fact that the weights are provided under Apache 2.0, the choice is whether to immediately invest in fine-tuning for distribution or first confirm the license and what is actually delivered; the premise for changing that choice is the files actually contained in the repository and the model card terms. The standard: check the license file and any additional terms of use, register the license metadata in your internal model inventory, and do not invest in fine-tuning for external distribution before legal confirmation.

Limits and open questions

The benchmark scores and the roughly 32 million US dollar cost in the report are all company self-reported; the report does not say whether third-party reproduction exists, and gives no evaluation configuration details. "Best US open model" is a claim about the single TerminalBench 2.1 benchmark only; it does not contradict the three lower results the report lists afterward, but it cannot be treated as an overall capability conclusion. The roughly 130GB VRAM estimate in this article is an arithmetic inference from 65 billion parameters at two bytes per parameter in BF16, not source data; the source gives no minimum VRAM, context length, quantized versions or inference framework support. Ripple-65B Preview is currently not in the KimiSeek model catalog; the source reports only weights on Hugging Face and does not mention a hosted API, so any reference to API availability concerns the vendor, not KimiSeek; if a vendor or third-party API appears later, availability and billing should be based on the provider's pages. Versions labeled Preview usually keep changing, so whether the weights will be maintained and whether training code and data are open should be checked against the model card and local testing.

Three minutes to your first working call

Create an account and an API key. Calls use your account balance. 30% off your first top-up. Top up ¥500, pay ¥350, receive ¥500 in credit. Top up ¥1000, pay ¥850, receive ¥1000 in credit You can still start with a ¥10 trial top-up.

Sources and editorial process

Ripple-65B Preview 发布:完全 Apache 2.0 开源的美国模型,智能体编程成绩强劲

AI assisted the analysis and translation of the cited sources, followed by automated quality checks. Recommendations are not test results. Consult the current catalog for availability and pricing.