Ricursive says its open Ripple-65B leads US open models at coding and fits a DGX Spark

The chip-design startup says the whole model cost $32 million to build, a tenth of the last record holder's.

By Think Facility · November 6, 2026 · Open weights Local models

Ricursive Intelligence released its first model, Ripple-65B Preview, on November 4, 2026, with the weights on Hugging Face under an Apache 2.0 license. It's a 65-billion-parameter model the company pretrained from scratch, and it's a general-purpose version of the model behind Ricursive's AI chip design platform.

Ricursive says it set a new high for US open-weight models on agentic coding. It also says the whole thing cost $32 million, start to finish, which it puts at a tenth of what the previous record holder cost.

Model: Ripple-65B Preview, 65B parameters, dense
License: Apache 2.0
TerminalBench 2.1: 80.52
Full R&D cost: $32M
Runs on: NVIDIA DGX Spark, per Ricursive

What Ricursive is claiming

The headline number is 80.52 on TerminalBench 2.1, a test of an AI agent working in a computer terminal. In Ricursive's model card that edges out Reflection's Beam 501B at 80.1 and leaves Meta's Muse Spark Open 816B at 76.4. Those rivals are mostly much bigger. Ripple is dense, though, so all 65 billion parameters work on every token.

A small gold-colored metal box with textured metal foam panels at the front, sitting on a wooden table — An NVIDIA DGX Spark, the desktop machine Ricursive says can run Ripple-65B. Photo: Daniel Lu ( User:dllu ), CC BY-SA 4.0, via Wikimedia Commons

It doesn't win everything in its own table. On DeepSWE 1.1, Ripple scores 44.3 against Muse Spark Open's 60.0, and on MCP Atlas, a tool-use test, it's 71.67 to 89.2. The card is candid about the mess underneath. It says replicating agentic benchmark scores has been challenging, and that it used its own evaluation runs wherever a lab hadn't published a number.

Anna Goldie posted the launch on X 22 minutes after the company did, and put the cost and the laptop claim up front.

Anna Goldie@annadgoldieNov 4, 1:46 pm ET

Thrilled to announce Ricursive's first model, Ripple-65B Preview, which we pretrained from scratch! We achieved a new U.S. open-weight SOTA on agentic coding at a tenth of the cost of the prior SOTA, and can run locally on a Macbook or DGX Spark. By adapting our model and training recipes to the hardware, we were able to achieve massive efficiency gains equivalent to training on 16x more GPUs. This result is a testament to the team's depth in both model and chip architecture.

View the post on X

How a small team did it for $32 million

The launch post says the $32 million covers everything (pretraining experiments, data, evals, debugging and the final runs), not just the last training job. The model card adds that the team grew from one person to five during development.

Most of the savings came from keeping things simple. Ricursive chose a plain dense design over mixture-of-experts, which it says kept its infrastructure manageable, and it reports 40% hardware utilization through training with no custom kernels. It didn't buy any post-training data either, building its own synthetic pipelines instead.

We reached a TerminalBench 2.1 score of 80.5 with only 1.4M rollouts, roughly 70x fewer rollouts than Reflection Beam, the prior SOTA.

From Ripple-65B Preview, a Frontier Open-Weight Model | Ricursive Intelligence

What it means if you want AI on your own machine

For people outside the industry, the part that matters is where the model can run. Ricursive says Ripple fits on consumer hardware such as NVIDIA's DGX Spark, a desktop box, and it ships a compressed NVFP4 checkpoint and a Colab notebook to get people started. Goldie and Azalia Mirhoseini both wrote on X that it runs on a MacBook too. Neither post says how much memory that takes.

That's worth checking before you try it. The full-precision files on Hugging Face add up to about 130 GB, which is far beyond most laptops, so local use means the smaller quantized versions. The Hugging Face page lists 30 of those already.

Running it locally means a law office or a clinic could keep its files off someone else's servers, and an Apache 2.0 license lets them use it commercially without paying Ricursive. I wouldn't treat it as a finished product, though. Ricursive calls this a Preview with relatively light post-training, aimed at coding.

How long the record lasts

Possibly a week. Thinking Machines released Inkling 2 Small on November 5, and its model card gives it 81.3% on TerminalBench 2.1 with the same Terminus harness. Its weights aren't public yet, so it doesn't count as open, but Thinking Machines says they're due next week.

Ripple-65B Preview is available now on Hugging Face, Together AI, GMI Cloud, OpenRouter and Fireworks.