Ricursive Intelligence’s Ripple-65B Preview claims coding lead among US open models

Ricursive Intelligence released Ripple-65B Preview on November 4, 2026 under an Apache 2.0 license, an open-weight dense Transformer the chip-design startup says leads all US open models on two coding benchmarks. Ripple-65B scored 80.52 on TerminalBench 2.1, edging Reflection Beam 501B at 80.1 despite carrying 65 billion parameters versus 501 billion, and posted 79.1 on SWE-bench Pro v2.0 Hard. Ricursive pretrained the model from scratch on 12 trillion tokens with a 1 million-token context window and 1.4 million reinforcement learning rollouts, at roughly $32 million in R&D; weights are on Hugging Face and the model is served through Together AI, GMI Cloud and OpenRouter.

Ricursive Intelligence’s Ripple-65B Preview claims coding lead among US open models The chip-design startup's first open-weight model edges past a far larger rival on TerminalBench 2.1 and posts 79.1 on SWE-bench Pro v2.0 Hard Ricursive Intelligence has released Ripple https://cryptobriefing.com/markets/ripple/ -65B Preview, an open-weight model that the company says leads all US open models on two demanding coding benchmarks. The startup is best known for using AI to design chips. Now it is shipping a general-purpose coding model, and it is putting up numbers against models many times its size. The release landed on November 4, 2026, under an Apache 2.0 license. That is about as permissive as licenses get: developers can download it, modify it, and use it commercially without asking anyone’s permission. The benchmark numbers Ripple-65B Preview scored 80.52 on TerminalBench 2.1, which measures how well a model writes and executes code inside a terminal environment. That score puts it ahead of Reflection Beam 501B, which sits at 80.1. The margin is less than half a point. The more interesting detail is size. Ripple has 65 billion parameters, while its closest rival carries 501 billion. The second benchmark is SWE-bench Pro v2.0 Hard, which covers difficult software engineering tasks. Ripple scored 79.1 there. The model is not a clean sweep, though. Ricursive’s own results show it trailing larger models on other knowledge and reasoning assessments. AI, tech, and the markets they move—in one daily briefing. Daily. Free. Join 34,000+ readers across crypto, finance, and policy. What’s under the hood Ripple-65B is a dense Transformer with 80 layers. “Dense” means every parameter is active for every query, unlike mixture-of-experts designs that only wake up part of the network at a time. Ricursive pretrained the model from scratch on 12 trillion tokens. It did not start from someone else’s base model and fine-tune it, which is the shortcut many smaller labs take. The context window stretches up to 1 million tokens. Post-training was comparatively lean. The company used only 1.4 million reinforcement learning rollouts, the trial-and-error phase where a model attempts tasks and gets rewarded for good answers. Ricursive puts total research and development costs at approximately $32 million, a figure it frames as significantly lower than what the prior open-weight leader cost to build. On hardware, the model runs on NVIDIA https://cryptobriefing.com/markets/nvidia/ datacenter GPUs. It has also been run on experimental setups such as DGX Spark https://cryptobriefing.com/markets/spark-2/ , NVIDIA’s compact desktop system. Where to get it The weights are available now on Hugging Face. For teams that would rather not host a 65-billion-parameter model themselves, Ripple is also served through API partners Together AI, GMI Cloud, and OpenRouter. Fireworks support is expected soon. Who is Ricursive Intelligence Ricursive was founded in late 2025 by AI researchers Anna Goldie and Azalia Mirhoseini. The company builds models that help generate more efficient designs for silicon chips. A $300 million Series A pushed the company’s valuation to $4 billion. The coding model is not a detour from that mission. Chip design is, at its core, a software-heavy discipline, and Ripple-65B is positioned to support Ricursive’s work on AI-accelerated chip design. Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy https://cryptobriefing.com/editorial-policy/ .