Ripple-65B is an open-weight coding model trained for $32 million

A small lab says it trained a top-tier open-weight coding model for a fraction of what the leaders spend. Ricursive Intelligence, a chip-design lab, released Ripple-65B Preview on 4 November 2026. The model is open-weight on Hugging Face and on four hosting services. Ricursive says the full training cost was $32 million.

Ricursive built the model to run on its own chip-design platform. The public release is a general-purpose version of the same model. It is pretrained on 12 trillion tokens and supports up to 1 million tokens of context.

The benchmark Ricursive reports

Ricursive reports a score of 80.52 on TerminalBench 2.1, a benchmark for agents that use a shell. By Ricursive's count, that is the top US open-weight score. Much larger models from other labs score lower on the same test, by Ricursive's list.

Ricursive puts the others at these scores, with their parameter counts:

• Reflection Beam: 80.1, 501 billion
• Meta Muse Spark Open: 76.4, 816 billion
• Thinking Machines Inkling: 63.8, 975 billion
• NVIDIA Nemotron 3 Ultra: 56.4, 550 billion

The comparison is a vendor benchmark. Ricursive ran it on its own model. It quotes the other models' reported scores. A user who wants to confirm the lead will run the four models side by side on real tasks.

Why the shape of the model matters

Ripple-65B is a dense Transformer, not a mixture-of-experts. Ricursive picked dense on purpose, to cut the complexity of training and serving. The model is thin and deep: 5,120 d_model and 80 layers. The company says this shape fits well on NVIDIA DGX Spark. The Spark is a desktop box with high compute and low memory bandwidth.

That is the second interesting number here. Most top open-weight models need a server rack to serve. This 65B dense model can be served on a single high-end workstation. Latency stays low. The full 1M context fits.

What the training looked like

Ricursive says it wrote its own synthetic data pipeline and did not buy any post-training data. On the reinforcement learning stage, it reached its 80.5 score with 1.4 million rollouts. Reflection Beam, the earlier state-of-the-art, used around 70 times more rollouts for a similar score, by Ricursive's estimate.

The company compares its full $32 million cost to two earlier public figures from other labs. Sam Altman said in 2023 that GPT-4's training cost was over $100 million. Dario Amodei later said frontier models could cost up to $1 billion to train. Ricursive's $32 million is one tenth of the previous open-weight leader's cost, the company says, on its own best estimate.

Ricursive also builds on Allen AI's Olmo 3 recipe. The company credits open-source work from Prime Intellect, RadixArk and NVIDIA too. Its starting point was the Olmo ladder. Four generations of changes made its ladder about 16 times more compute-efficient than the baseline at 4B parameters.

Where to try it

Ripple-65B Preview is on Hugging Face, Together AI, GMI Cloud and OpenRouter, Ricursive says. Fireworks goes live later on the same day.

On Hacker News, the launch post reached 201 points in five hours. A buyer picking an open-weight coding model this week now has one more candidate. It also fits on a smaller box than most of its rivals.

Source

• Ripple-65B Preview: Achieving US Open-Weight SOTA on Agentic Coding, Ricursive Intelligence
• Hacker News discussion