Ricursive drops Ripple-65B: A dense model outcoding 900B MoE giants

Almost every major AI lab today is obsessed with Mixture-of-Experts (MoE) architectures, bundling hundreds of billions of parameters to keep active compute per token low.

Ricursive Intelligence, a lab founded by former Google researchers, decided to bet in the opposite direction. The team just open-sourced Ripple-65B Preview under an Apache 2.0 license, delivering a pure dense 65B Transformer model with 1M context that outperforms several frontier MoE models ten times its size.

A $32M bet built by five people

Ripple-65B was originally created to drive Ricursive's proprietary AI chip design platform. Ripple-65B Preview is the general-purpose variant of that internal engine.

What catches the eye first is the budget. While industry leaders talk about pretraining bills heading toward a billion dollars, Ricursive spent just $32M from scratch. That figure covers their entire R&D pipeline: model laddering, data mixture experiments, debugging, synthetic data generation, and the final hero training run.

Even more striking is the team footprint. Ricursive scaled from 1 to just 5 people during development. They pre-trained the model on 12 trillion tokens from scratch, holding a steady 40% Model Flops Utilization (MFU) in BF16 using clean sharding without custom kernels.

How a 65B dense model beat 900B MoE giants

Rather than taking on the routing overhead and multi-node communication complexity of MoE, Ricursive went with a thin-and-deep dense layout. The model packs 80 layers with a relatively narrow hidden dimension of 5120, taking cues from Gemma 3N-style designs to maintain training stability at scale.

They paired this layout with gated attention, value residuals, FFN activation clipping, and global NoPE. On the post-training side, they avoided buying commercial datasets and instead built tight synthetic data filtering and high-efficiency reinforcement learning. With only 1.4M rollouts—about 70 times fewer than the prior open-weight leader Reflection Beam—the model achieved an aggressive climb in capabilities.

On the agentic coding benchmark TerminalBench 2.1, Ripple-65B Preview scored 80.52. That beats significantly larger American open-weight systems, including 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.

The base and instruct weights are out on Hugging Face alongside an NVFP4 quantized checkpoint, meaning developers can run it on hardware like the NVIDIA DGX Spark. On two B200 GPUs, it clocks token generation speeds of 250 tokens per second.

Why it matters

For engineers looking to host their own coding agent, Ripple-65B is a breath of fresh air. Deploying a 500B to 900B MoE model requires massive GPU clusters simply to keep the unactivated weights in VRAM. A 65B dense model dramatically cuts down memory capacity demands and infrastructure headaches while matching top-tier agentic performance.

For the wider industry, Ricursive just proved that dense architectures still have serious runway. Brute-forcing massive parameter counts often masks poor data recipes and noisy training pipelines. A small, disciplined team optimizing data ladders and reinforcement learning efficiency can still outmaneuver tech giants.

We think Ricursive's approach is remarkably sharp, demonstrating how tightly integrated hardware-software co-design can punch above its weight. The caveat is that a 65B dense model fires all its parameters on every single token, so high-concurrency cloud hosting will still burn more raw compute per request than sparse MoE setups. But for local deployments and heavy reasoning tasks, it is one of the most compelling open models released this year.