AI News
All stories
Reported

Recursive Intelligence releases open-weights Ripple 65B preview model

Commentator Matthew Berman says the 65B dense Ripple model supports a 1M-token context and scores 80.5% on Terminal Bench 2.1 among open-weights models.

Open Source · · Updated · 3 min read

Recursive Intelligence, which its own website spells Ricursive Intelligence, has released Ripple-65B Preview, a 65-billion-parameter dense open-weight language model. The company says it sets a new US open-weight state of the art on agentic coding, with a score of 80.52 on TerminalBench 2.1. The weights are available on Hugging Face, Together AI, GMI Cloud, OpenRouter and Fireworks. The company describes the model as a general-purpose variant of a proprietary model that powers its AI chip design platform. It says the model was pretrained on 12 trillion tokens with support for up to 1M tokens of context.

The architecture is one of the design choices the company highlights. The company notes that other labs often use mixture-of-experts designs, which activate only part of the network for each token. Ripple instead uses a dense Transformer, so every parameter takes part in every token. The company argues this reduces the infrastructure complexity of pretraining and posttraining while still allowing strong reasoning. It says the model kept a 40% MFU, a measure of how much of the hardware's peak throughput training used, even with simple sharding, no custom kernels and no quantization. The network is what the company calls thin and deep: a model width of 5120 across 80 layers, with overprovisioned attention heads and a shift in the mix of sliding-window and global attention layers.

Matthew Berman, a YouTube commentator who covers AI releases, reported the benchmark and framed it against other open-weight models only. His chart placed Ripple at 80.5% on Terminal Bench 2.1, ahead of Reflection Beam at 501B parameters (80.1), Meta Muse Spark Open at 816B (76.4), Thinking Machines Inkling at 975B (63.8) and NVIDIA Nemotron 3 Ultra at 550B (56.4). The company's own page makes the same comparison and stresses that the competing models are significantly larger.

The company's cost claim is the other headline figure. It says the full research and development cost came to $32M. That includes pretraining ladder development, data mixing, evaluations, debugging, synthetic data generation, posttraining and the final pretraining run. The company calls this one-tenth the cost of the previous open-weight state of the art on agentic coding, as estimated to the best of its knowledge. A footnote cites reports that Reflection committed $150M per month for compute at SpaceX Colossus and that its Beam model took 4 weeks to pretrain and 4 weeks to posttrain.

The company says its reinforcement learning reached the 80.5 score with only 1.4M rollouts, roughly 70 times fewer than Beam. It says it bought no posttraining data and relied instead on its own synthetic data pipelines and quality filtering. Its pretraining work started from the model ladder that the Allen Institute for AI published for Olmo 3. The company describes a series of ladder versions in which it tested changes at small scale. One variant was rejected because it did not improve at 4B scale. The company reports that the final changes improved compute efficiency on code by up to 16x at 4B scale, as measured by what it calls dolma_stack CE.

Berman says the model is small enough to run locally, on a MacBook or an NVIDIA DGX Spark, and he quotes prices of 20 cents per million input tokens and 60 cents per million output tokens. The company's page names the DGX Spark as an example of a local device that can run the model efficiently, but the page doesn't list prices, so those figures come from Berman alone. Berman also presented Ripple as part of a trend of US companies releasing small and mid-sized open-weight models. He pointed to Thinking Machines' Inkling 2 Small, which he said scored 31 on the Artificial Analysis Intelligence Index.

What changed

Recursive Intelligence has released Ripple-65B Preview, a 65-billion-parameter dense open-weight language model.

Type: New capability · Applies to: Ripple-65B Preview

Why it matters

A small open-weights model with long context and low quoted prices gives builders a self-hostable option for agent workloads.

How the story developed

  1. Adds the maker's claim that the dense model is the top US open-weight model for agentic coding, cited from its own post.

Original source

Where we saw it

2 sources

Forward Future (Matthew Berman), Latent Space

Recursive Intelligence

AI wrote this summary from the sources listed, and nobody reviewed it individually before publication. Check the original source before you rely on it. How it works