Ricursive Releases Open Ripple-65B Coding Model It Says Cost $32 Million to Build

The model is a dense Transformer with 65 billion parameters, pretrained on 12 trillion tokens and able to handle context of up to 1 million tokens, the company said on its website. It was released this week under the Apache 2.0 licence and is available through Hugging Face, Together AI, GMI Cloud, OpenRouter and Fireworks.

The release matters for two reasons. It comes from a company whose main business is chip design rather than model building, and its claimed cost is a fraction of what larger labs spend. Ricursive said the $32 million covers all research and development, including experiments, data work, evaluations and the final training runs, and estimated that this is "one-tenth the cost of the previous open-weight SOTA on agentic coding". That figure is the company's own estimate.

On TerminalBench 2.1, a test of coding agents working in a command-line environment, Ricursive reported a score of 80.52. It lists Reflection's 501-billion-parameter Beam at 80.1, Meta's 816-billion-parameter Muse Spark Open at 76.4, Thinking Machines' 975-billion-parameter Inkling at 63.8 and Nvidia's Nemotron 3 Ultra at 56.4. All of these results are self-reported, and The Decoder noted that Ripple's harness was partly Terminus 2 and partly OpenCode.

Other benchmarks are less flattering. According to figures compiled by The Decoder, Ripple scored 79.1% on SWE-Bench Pro v2.0 Hard against 86.3% for Thinking Machines' newly released Inkling 2 Small, 44.3% on DeepSWE v1.1 against 72.1%, and 71.67% on MCP Atlas against about 85%. The Decoder's assessment was that Ripple is likely easier to run on a user's own hardware.

That is where Ricursive is pitching it. The company said the model is small enough to run locally on an Nvidia DGX Spark desktop system, and that it chose a dense design over the mixture-of-experts approach used by most large open models to keep token generation fast. The model was built to train and serve efficiently on Nvidia hardware.

Ricursive described Ripple as a general-purpose variant of the proprietary model that powers its AI chip design platform. The company said post-training used no purchased data, relying instead on its own synthetic data pipelines and quality filters, and that it reached its TerminalBench score with 1.4 million reinforcement learning rollouts, about 70 times fewer than Reflection used for Beam. It also said it kept model flops utilisation at about 40% through training without custom kernels or quantisation.

The company, founded by former Google DeepMind researchers Anna Goldie and Azalia Mirhoseini, says its aim is "closing the loop between models and the hardware that fuels them". It labels this release a preview and says its next step is to scale the approach towards frontier-sized models.