Ricursive Intelligence, a startup founded by former Google researchers that builds AI for chip design, has released an open-weight model called Ripple-65B Preview. It is much smaller than most headline open models, yet the company says it leads U.S. open-weight models on agentic coding. Here is what is confirmed and what still needs independent testing.

Coffee summary

  • FACT: Ripple-65B Preview is a 65-billion-parameter dense model, trained from scratch on 12 trillion tokens, with up to a 1M-token context.
  • FACT: The weights are on Hugging Face under the Apache 2.0 license, which allows commercial use and modification.
  • CLAIM: Ricursive reports 80.52 on TerminalBench 2.1, ahead of much larger U.S. open models such as Reflection Beam 501B (80.1).
  • CLAIM: The company says its full R&D cost was only $32 million.

What happened

Ricursive trained Ripple to power its own chip design platform and released a general-purpose version. It is available on Hugging Face, Together AI, GMI Cloud, and OpenRouter, with Fireworks listed as coming online.

Unlike many recent large models that use a “mixture of experts” design, where only part of the model runs for each word, Ripple is a dense model: all 65 billion parameters are used every time. Ricursive says that choice kept training simpler and gives fast, low-latency output. The company also says the model is small enough to run on local hardware such as NVIDIA’s DGX Spark desktop system.

Why it matters

Most top open models are now enormous. Kimi K3.1 Pro has 2.8 trillion parameters and Mistral Large 4 about 1 trillion, so they need data center hardware. A 65B model with strong coding skills is far easier for a company or a serious hobbyist to host privately. The Apache 2.0 license also has none of the revenue limits or display rules some other open models carry.

Who should care

  • Teams that want a self-hosted coding assistant for privacy or cost reasons.
  • Developers building agents who want a permissive license.
  • Researchers interested in efficient training, since Ricursive shared many architecture details.

Limitations

This is a “Preview,” and all benchmark scores and the $32 million cost are the company’s own figures. TerminalBench measures terminal-based coding tasks, not every kind of work. A 65B dense model still needs a large-memory GPU or a high-end workstation; it will not run on a normal laptop at full precision. Check independent evaluations before switching a production workflow.

What to do next

  1. Try it cheaply first through OpenRouter or Together AI on a few real tasks from your own codebase.
  2. Compare it with the model you use now on the same prompts, not just on benchmarks.
  3. If results are good, review the hardware needs and the Apache 2.0 notice requirements before self-hosting.

AIImpish Take

Ripple-65B Preview is a promising sign that strong coding models do not have to be giant. The claims are impressive but self-reported. Test it on your own work this week; if it holds up, it could be one of the easiest serious coding models to run privately.

Related on AIImpish

Sources: Ricursive Intelligence, Introducing Ripple-65B Preview; Hugging Face, ricursiveAI/Ripple-65B-Preview; VentureBeat.