Ripple-65B Preview

Ripple-65B Preview is a 65B dense open-weight language model released on 2026-11-04 by chip-design lab Ricursive Intelligence (co-founders include Anna Goldie and Azalia Mirhoseini). Ricursive describes it as a general-purpose variant of the proprietary model behind its AI chip-design platform: pretrained on about 12 trillion tokens, advertising up to roughly 1M context, and self-reporting 80.52 on TerminalBench 2.1—ahead of much larger U.S. open-weight peers in their table (e.g. Reflection Beam 501B, Meta Muse Spark Open 816B).

Weights ship under Apache 2.0 on Hugging Face (ricursiveAI/Ripple-65B-Preview and companions), with hosting on Together AI, GMI Cloud, and OpenRouter, plus NVFP4 / Base / DFlash2 variants. The lab highlights a dense (non-MoE) design, ~$32M full R&D cost, and efficient local runs on MacBook Pro / DGX Spark-class hardware.

Problem

U.S. open-weight models have lagged closed/overseas MoE peers on agentic coding. Ripple aims to push TerminalBench-class work into a locally deployable 65B dense envelope via hardware–software co-design.

Differentiation

• Dense 65B instead of giant MoE—more predictable latency/serving.
• Full-stack chip+model story and a ~1/10 prior open SOTA R&D-cost narrative.
• Not the unrelated “Recursive”/Flint company—this is Ricursive Intelligence.

Limits

• Headline benchmarks are vendor-reported; third-party boards may lag.
• “Up to 1M” on the card vs ~16K on some hosted endpoints—weights ≠ API.
• Preview quality; validate long-context and tool use yourself.

Who it’s for

Teams wanting commercially usable Apache weights for terminal/repo agents who can afford 65B local or hosted inference.

Get started

1. Read ricursive.com/ripple and HF ricursiveAI/Ripple-65B-Preview.
2. Try Together / OpenRouter / GMI, or self-host BF16/NVFP4.
3. Replay your own TerminalBench-like tasks before swapping coders.