Ricursive 65B Open Model Beats 501B Giant in Coding Benchmarks

A chip design startup dropped a 65B open-weight model that outperforms a 501B giant on terminal coding tasks.

Models · Source: VentureBeat

What happened

Ricursive Intelligence has officially released Ripple-65B Preview to the public. It is an open-weight model distributed under an Apache 2.0 license. This means developers can download, modify, and use it for commercial applications without asking for permission. The model is a dense Transformer with 80 layers. Unlike mixture-of-experts designs that only activate parts of the network, every parameter in Ripple is active for every query. Ricursive pretrained the model entirely from scratch using 12 trillion tokens instead of taking the shortcut of fine-tuning an existing base model.

The performance numbers are turning heads. Ripple-65B Preview claims the lead among US open models on two demanding coding benchmarks. It scored 80.52 on TerminalBench 2.1, a test that measures how well a model writes and executes code inside a terminal environment. This score puts it slightly ahead of Reflection Beam 501B, which scored 80.1. The margin is small, but Ripple achieved this with a fraction of the size. It also posted a strong score of 79.1 on SWE-bench Pro v2.0 Hard, an assessment covering difficult software engineering tasks.

The startup kept post-training remarkably lean. They used only 1.4 million reinforcement learning rollouts to teach the model how to complete tasks. Ricursive reports their total research and development costs at approximately 32 million dollars. The model features a massive context window stretching up to 1 million tokens. Developers can grab the weights right now on Hugging Face or access the model through API partners like Together AI, GMI Cloud, and OpenRouter.

Key facts

Why it matters

This release completely changes how founders should think about model size and capability. A 65-billion parameter model beating a 501-billion parameter giant proves that sheer scale is no longer the ultimate advantage. Smart training data and focused architecture matter more. Builders now have access to a highly capable coding model that does not require a massive server farm to operate. It runs on standard NVIDIA datacenter GPUs and has even been tested on compact desktop systems like the DGX Spark. This lowers the barrier to entry for building complex software agents.

The second-order effect is the rise of vertical AI companies producing general-purpose foundation models. Ricursive is a chip design startup valued at 4 billion dollars after a 300 million dollar Series A. They built Ripple-65B to support their own AI-accelerated chip design work. Because chip design is a software-heavy discipline, solving their internal problems resulted in a world-class coding model as a byproduct. We will likely see more specialized startups releasing powerful open models that rival the output of dedicated AI research labs.

For builders

Run elite coding agents locally

Ripple-65B fits on compact experimental setups like the DGX Spark desktop system. Founders can save massive cloud computing costs by running complex software engineering tasks entirely in-house.

Commercialize without licensing headaches

The Apache 2.0 license is as permissive as it gets for open-weight models. You can modify the code and sell products built on top of it without paying royalties or asking permission.

Leverage massive context windows

The model supports a context window of up to 1 million tokens. Engineers can dump entire codebases or extensive documentation into the prompt for deep debugging and automated refactoring.

My take

Stop obsessing over parameter counts. I look at these numbers and see that smart training beats bloated models every single time. If a chip startup can build a world-class coding model as a side project for 32 million dollars, the massive AI labs should be very nervous.

Original reporting: VentureBeat. This is my rewrite and opinion.

More AI news for builders