Ricursive Intelligence, a startup that builds AI for chip design, has released a general-purpose language model called Ripple-65B Preview, and says it cost $32 million to develop. The company announced the release on 4 November in a launch post and a Hugging Face model card. If the figure holds up, it undercuts the nine-figure budgets that have become the reference point for frontier training.

Ripple is a dense Transformer with 65 billion parameters, trained from zero on 12 trillion tokens, most of it English prose and source code. It handles up to 1 million tokens of context and ships under the Apache 2.0 license, which allows commercial use. The release includes an instruction-tuned version in BF16 and NVFP4 formats plus a BF16 base model. Ricursive built the original to run its own chip design platform, and Ripple is a general-purpose sibling of that proprietary system.

The headline result is a TerminalBench 2.1 score of 80.52, on a test of agents working in a command line. Ricursive calls that the best score among U.S. open-weight models. It lists Reflection Beam 501B at 80.1, Meta’s Muse Spark Open 816B at 76.4, Inkling 975B from Thinking Machines at 63.8, and NVIDIA’s Nemotron 3 Ultra 550B at 56.4. Those rivals use mixture-of-experts designs that activate 23 billion to 55 billion parameters per token, so the comparison pits a small dense model against much larger ones. Ricursive describes its table as “a best-effort attempt at a fair comparison” and says Reflection Beam falls to 70.8 under the official Terminus 2 harness.

The model card does not claim a clean sweep. On the SWE-bench Verified coding test, Ripple lands at 76.5 against 81.6 for Reflection Beam and 77.6 for Inkling. On MCP Atlas it posts 71.67 against 89.2 for Muse Spark Open, and on SciCode it gets 39.4 against that same model’s 58.5. The harder SWE-bench Pro v2.0 variant goes its way, 79.1 to Reflection Beam’s 76.5. Every one of these numbers comes from Ricursive, and the launch post cites no independent replication.

The $32 million claim deserves the same caution. Ricursive says it covers everything: small-scale ladder experiments, data mixing, evaluations, debugging, synthetic data, and the last big training stages. The company estimates that is a tenth of what the previous open-weight leader on agentic coding cost, and sets it beside Sam Altman’s statement that GPT-4 cost more than $100 million. Nobody outside the company has audited the total.

The efficiency story is where the details get interesting. A team that grew from one to five people got to its TerminalBench result on 1.4 million reinforcement learning rollouts, about 70 times fewer than Ricursive says Reflection Beam used. It purchased no data for post-training. It reports 40 percent hardware utilization using plain sharding, and says it needed neither hand-written kernels nor quantization, and credits a head start from the Olmo 3 recipe published by the Allen Institute for AI.

For builders the practical draw is size. Ricursive says Ripple is small enough for consumer-class machines like NVIDIA’s DGX Spark, and that two B200 GPUs push out 250 tokens a second at a batch size of 8. Hosted access is live on Fireworks, OpenRouter, Together AI, GMI Cloud, and Hugging Face, so trying it costs an API call.

Ripple’s results are lopsided, strong on terminal work and weaker on tool-calling and science. Teams choosing an open model for coding agents can test it against their own tasks this week, and the first independent TerminalBench runs will show whether the 80.52 survives outside Ricursive’s harness.

Reported by Ricursive Intelligence in its launch post and Hugging Face model card for Ripple-65B Preview, 4 November 2026.