Ripple-65B Preview Claims 80.52 on TerminalBench 2.1 With a 65B Dense Model

Sanket Chaukiyal

November 17, 2026

TL;DR

  • Ricursive Intelligence released Ripple-65B Preview, a 65B dense Transformer pretrained on 12 trillion tokens, with support for up to 1M context.
  • Its model card reports 80.52 on TerminalBench 2.1, ahead of Reflection Beam 501B (80.1), Meta Muse Spark Open 816B (76.4), Thinking Machines Inkling 975B (63.8) and NVIDIA Nemotron 3 Ultra 550B (56.4).
  • Every score is self-reported, and the authors themselves flag replication challenges that depend on the test harness, such as Terminus 2.
  • The repository hosts a BF16 base model under the Apache 2.0 license.

A 65B Dense Model Crashes the Agentic Coding Table

Ricursive Intelligence has put out Ripple-65B Preview on Hugging Face. It’s a 65-billion-parameter dense Transformer, pretrained on 12 trillion tokens, with context support stretching to 1M.

The headline is the benchmark. The model card says Ripple-65B Preview scores 80.52 on TerminalBench 2.1, “surpassing significantly-larger US open-weight models” and then lists them: Reflection Beam 501B at 80.1, Meta Muse Spark Open 816B at 76.4, Thinking Machines Inkling 975B at 63.8, and NVIDIA Nemotron 3 Ultra 550B at 56.4.

Size is the story here. Inkling is about 15 times larger by parameter count, and even the closest rival, Beam, is nearly eight times bigger. The weights are a BF16 base model released under Apache 2.0, so anyone can pull them down and poke at them.

One caveat sits right next to the headline. These numbers come from the model’s own creators, not an outside lab.

Why a 0.42-Point Lead Is Both Tiny and Interesting

Start with the gap. Ripple scores 80.52 and Reflection Beam 501B scores 80.1. That’s a difference of 0.42 points, on a self-reported test, where the authors admit results shift with the harness. I wouldn’t hang a procurement decision on it.

But the gap to everyone else is the real signal. Muse Spark Open 816B trails by about four points. Inkling and Nemotron trail by far more. If the card’s numbers survive contact with other people’s setups, a 65B dense model isn’t merely keeping up with much larger systems on agentic coding. It’s sitting at the front of the pack.

Picture a marathoner who shows up with a lighter pack and finishes a shoe-length ahead of the leader, except the timing mats were laid out by his own crew. The finish might be real. You’d still want a second set of mats before you called it a record.

Why does the dense-versus-MoE framing matter? The verified pitch is practical: top-tier agentic coding from a 65B dense model could run on smaller datacenter nodes and on consumer-class hardware such as NVIDIA DGX Spark, without the infrastructure overhead that big mixture-of-experts systems drag along. That’s the claim, and nobody has independently tested it yet. Still, it’s the right question to ask.

And the comparison cuts at the competitors. Beam, Muse Spark Open, Inkling and Nemotron all carry far bigger parameter counts, and on this one benchmark that size didn’t buy them a lead. Parameter count as a proxy for agentic capability looks shakier if Ripple’s result holds.

I’ve read plenty of model cards that shine until someone else runs them. My bet is the harness footnote will matter more than the 80.52. Credit to the authors for putting it in writing.

Olmo 3 Roots and a Team That Grew From One to Five

The build story is almost as unusual as the score. According to the card, the team scaled from 1 to 5 people and trained the model from scratch. They held 40% model FLOPs utilization (MFU) during pretraining, without MoE or quantization.

The architecture didn’t start from a blank page. The model began from an initial Olmo 3 baseline, then moved to a Gemma 3N-like design. The card lists a narrow d_model of 5120 across 80 layers, FFN activation clipping, gated attention, global NoPE and value residuals.

Thin and deep is the pattern. Ricursive Intelligence says it chose this modernized dense layout to keep infrastructure simple while matching or beating larger MoE models on agentic coding. That’s the stated reasoning, and the card doesn’t offer an ablation in the material we’ve verified, so treat the causal story as the authors’ own.

The label matters too. This is a Preview, and the repository hosts the base model. Anything built on top, including the post-training that agentic work usually leans on, isn’t something the verified material describes.

Three Checks That Will Settle the Claim

First, watch for independent TerminalBench 2.1 runs. The authors say replication across agentic evaluations is tricky and depends on harnesses like Terminus 2. If outside teams land near 80 with their own setups, the story gets much stronger. If scores scatter, the 0.42-point lead over Beam evaporates fast.

Second, look for hardware reports. The pitch about smaller datacenter nodes and DGX Spark-class machines is the practical payoff, and it’s only testable by people who actually load the weights. Real throughput and memory numbers from outside users would tell you more than another benchmark table.

Third, see what Ricursive Intelligence does after Preview. A full release, further evaluations beyond TerminalBench 2.1, or documentation of the training recipe would show whether this was a one-benchmark spike or a durable result. Would a lab with a five-person team keep that pace? That’s worth tracking.

Editor's Note

What caught my eye wasn't the score. It was the team size. A group that grew from one person to five, trained from scratch, and still lists four much bigger names behind it on one benchmark. I'm not ready to believe the 0.42-point lead over Beam, and the authors' own harness caveat tells me they aren't either. I'm watching for someone outside Ricursive to run it on a different setup and post the number.

– Sanket Chaukiyal, founder, SmartChunks

FAQ

What is Ripple-65B Preview?

It's a 65B parameter dense Transformer from Ricursive Intelligence, pretrained on 12 trillion tokens with support for up to 1M context. The Hugging Face repository hosts the BF16 base model under the Apache 2.0 license.

What did it score on TerminalBench 2.1?

The model card reports 80.52. It lists Reflection Beam 501B at 80.1, Meta Muse Spark Open 816B at 76.4, Thinking Machines Inkling 975B at 63.8 and NVIDIA Nemotron 3 Ultra 550B at 56.4.

Are the benchmark results independently verified?

No. The scores are self-reported by the model's creators. The authors also note that replication across agentic evaluations can vary depending on test harnesses such as Terminus 2.

How was Ripple-65B trained?

A team that scaled from 1 to 5 people trained it from scratch, holding 40% MFU without MoE or quantization in pretraining. The design began from an Olmo 3 baseline and moved to a Gemma 3N-like architecture with 80 layers and a d_model of 5120.

Sanket Chaukiyal — Editor at Smart Chunks

Sanket Chaukiyal

Technology editor • 12+ years in editorial

Sanket is the founder and editor of Smart Chunks. He spent over six years at Autocar India (Haymarket SAC Publishing) as Sub Editor and Senior Copy Editor, and later served as Account Director (Content) at Rite Knowledge Labs. He holds a Master's in Media and Communication from the Symbiosis Institute of Media and Communication.

All articles → • LinkedIn