Ricursive Intelligence, a startup that builds AI systems for chip design, has released Ripple-65B Preview, an open-weight language model aimed mainly at agentic coding. The weights are on Hugging Face under the Apache 2.0 license, which allows commercial use and modification, and the model is also served by Together AI, GMI Cloud and OpenRouter, with Fireworks support promised. VentureBeat, which covered the launch on November 5, 2026, describes the founders as former Google researchers.
The model is a general-purpose version of the proprietary system behind Ricursive's chip design platform. It is a dense Transformer with 65 billion parameters, pretrained from scratch on 12 trillion tokens of mostly English text and code, and supports context windows of up to 1 million tokens. The company chose a dense design rather than the mixture-of-experts approach most labs now use, arguing that it kept training and serving simpler. The release includes an instruction-tuned model in BF16 and NVFP4 formats and a BF16 base model.
Ricursive's headline claim is a score of 80.52 on Terminal-Bench 2.1, just ahead of Reflection's Beam 501B at 80.1 and well above Meta's Muse Spark Open, Thinking Machines' Inkling and NVIDIA's Nemotron 3 Ultra, all of which are far larger by total parameter count. On SWE-bench Pro v2.0 Hard it reports 79.1 against 76.5 for Beam. The model card shows a less uniform picture elsewhere: Ripple trails Beam on SWE-bench Verified, scores well below Muse Spark Open on DeepSWE 1.1 and MCP Atlas, and lags several rivals on knowledge tests such as GPQA Diamond. Ricursive also notes that replicating agentic benchmark scores is difficult and calls its comparison a best-effort attempt.
Cost is the other part of the pitch. Ricursive says the full research and development bill, from early scaling experiments to the final training runs, was $32 million, which it estimates at about one tenth of what the previous open-weight leader in agentic coding required. It reached its Terminal-Bench result with 1.4 million reinforcement learning rollouts, roughly 70 times fewer than Beam by its account, and says the team grew from one person to five during development.
The architecture is narrow and deep, with 80 layers and a model width of 5,120, plus changes such as gated attention and value residuals. Ricursive started from the training ladder published by the Allen Institute for AI for Olmo 3 and says its own adjustments made code training up to 16 times more compute-efficient at small scale. Because the model is relatively compact, the company says the NVFP4 version runs on a single NVIDIA DGX Spark, and it reports about 250 tokens per second on two B200 GPUs before any quantization.
Every benchmark and cost figure here comes from Ricursive; no independent evaluation was available at release. The label Preview and what the company calls relatively light post-training suggest the model will change, and Ricursive has not said when a full version or its internal chip design variant might follow.