Ripple-65B Beats 975B Rivals on Agentic Coding for Just $32M

The small model that outruns the giants

On paper, Ripple-65B Preview should not be in this fight. It is, and it wins.

Ricursive trained Ripple-65B Preview from scratch on 12 trillion tokens with support for up to a 1M-token context window. On TerminalBench 2.1, an agentic coding benchmark, it posts 80.52 — the company's claimed US open-weight state of the art.

The comparison list is the interesting part. Reflection Beam 501B lands at 80.1. Meta Muse Spark Open 816B sits at 76.4. Thinking Machines Inkling 975B scores 63.8, and NVIDIA Nemotron 3 Ultra 550B trails at 56.4. A 65B dense model is holding its own against systems many times larger.

Ricursive credits a modernized dense Transformer rather than the Mixture-of-Experts approach most labs now favor. Narrow width (d_model 5120), 80 layers, extra attention heads, gated attention, global NoPE and value residuals round out the architecture. The payoff was a steady 40% MFU through pretraining and posttraining with simple sharding, no custom kernels and no quantization.

Ripple-65B Preview: size, context and availability

The $32M number is the real headline

Training frontier-adjacent models is supposed to cost a fortune. Ricursive says it didn't.

The company puts its full R&D bill for Ripple-65B Preview at $32M — and that figure covers everything from scratch: pretraining ladder development, data mixing, evals, debugging, synthetic data generation, posttraining recipe work, and the final hero run. By its own estimate, that is about one-tenth the cost of the previous open-weight SOTA on agentic coding.

For context, Sam Altman said in 2023 that GPT-4 training exceeded $100M, and Dario Amodei has projected pretraining costs of $10-100B for frontier models by 2026-2027. A $32M all-in figure lands in a different universe.

Ricursive did not buy posttraining data. It built its own synthetic pipelines, quality filtering and coverage metrics, then iterated on supervised fine-tuning and reinforcement learning recipes with compute efficiency as the goal. The result was a steep RL climb from a strong SFT checkpoint using only 1.4M rollouts — roughly 70x fewer than Reflection Beam needed.

Where you can actually run it

Ripple-65B Preview is not a waitlist situation.

The model is available now on Hugging Face, Together AI, GMI Cloud and OpenRouter, with Fireworks going live later today. Ricursive describes it as a general-purpose variant of the proprietary model that powers its internal AI chip design platform.

That last detail matters more than the benchmark. Ricursive is a full-stack lab trying to close the loop between models and the silicon they run on, and Ripple is the public face of that effort. The company says the dense architecture and small parameter envelope let the model run efficiently on local devices like NVIDIA DGX Spark, which have high compute but low memory bandwidth compared with datacenter accelerators.

What it means for you

For developers in the US, UK and EU, Ripple-65B Preview is accessible today through Hugging Face and the listed inference providers, so there is no regional gate to worry about. Pricing per token will depend on the provider you pick, and Fireworks has not yet published its rates.

The bigger practical question is hardware. Ricursive says the model runs efficiently on a DGX Spark, which puts a capable agentic coding model on a desk rather than behind an API. That is a meaningful shift for teams that want local inference without datacenter bills.

A 65B model matching 500B-plus rivals on agentic coding, built for the price of a modest startup round, is the kind of result that makes lab budgets look bloated. Ricursive's real test comes next: whether the same recipe scales, and whether independent benchmarks agree with its numbers.