x
Black Bar Banner 1
x

A+ A−
Ads by Markethive - View All
The Blog Feed
The Blog Feed
» Blog Feed
Write a New Blog Post
Search Blog Status
Most Viewed
Most Recent
Most Shared
Alphabetical
Markethive Blog (default)
My Blog Posts
Friends' Blogs
Blog Categories
All
Advertising
Blockchain & Cryptocurrency
Business Development
Diet & Weight Loss
Environmental
Health and Wellness
History and Culture
Home and Garden
Marketing
Mentoring & Training
Money & Finance
Other
Political
Prayer & Religion
Programming & Technical
Real Estate
Search Engine Optimization
Social Media
Spirituality
Sports & Recreation
Transport
Travel & Events
Website Design
Blogging Tools & Assets
My Blog Info
Members Subscribed to You
Blogs You Are Subscribed To
Website Widget
Wordpress Plugin
Subscribe for Greater Services
Subscribe to one of many subscriptions, each one includes the previous ones.. Unlock powerful tools, advance features, to build a powerful reach.

Ex-Google Pioneers Launch Ripple-65B Preview to Claim Open-Source Coding Lead 🚀

Posted by Simon Keighley on


Ex-Google Pioneers Launch Ripple-65B Preview to Claim Open-Source Coding Lead 🚀

Ex-Google Pioneers Launch Ripple-65B Preview to Claim Open-Source Coding Lead

The open-weight artificial intelligence landscape has just witnessed a remarkable shift. A lesser-known startup founded by former Google researchers has unveiled a new model that takes top honours on two demanding software engineering benchmarks, outperforming open-weight rivals that are up to fifteen times larger.

The model, named Ripple-65B Preview, comes from Ricursive Intelligence—a lab established by CEO Anna Goldie and CTO Azalia Mirhoseini, who previously helped lead Google’s groundbreaking AlphaChip initiative. Released under a permissive Apache 2.0 licence, Ripple-65B Preview sets a new benchmark for American open-source software development capabilities, as detailed in recent VentureBeat coverage.

 

Punching Above Its Weight on Agentic Coding

In artificial intelligence evaluation, agentic coding benchmarks test a model's ability to operate inside command-line terminals, execute multi-step tool calls, diagnose complex bugs, and write real-world software patch solutions.

On Terminal-Bench 2.1, Ripple-65B Preview scored 80.52, narrowly surpassing Reflection Beam 501B (80.1) and comfortable leaving behind Meta Muse Spark Open 816B (76.4), Thinking Machines Inkling 975B (63.8), and NVIDIA Nemotron 3 Ultra 550B (56.4). It achieved a similarly impressive 79.1 on SWE-bench Pro v2.0 Hard, establishing its position at the top of current U.S. open-weight agentic evaluation leaderboards.

What makes these results particularly striking is parameter efficiency. While competing frontier models rely on massive Mixture-of-Experts architectures spanning hundreds of billions of total parameters, Ripple-65B Preview achieves its results using a dense architecture with 65 billion active parameters. Full benchmark results and model card comparisons can be viewed directly on the Ripple launch announcement.

 

The Efficiency Breakthrough: Frontier Performance at One-Tenth the Cost

Building frontier AI models normally demands eye-watering budgets that stretch into hundreds of millions of dollars. Ricursive Intelligence took a different path, engineering and training Ripple-65B Preview from scratch for approximately $32 million in total research and development spend—roughly one-tenth the estimated cost of prior state-of-the-art open-weight models.

The team achieved this through two core innovations:

  1. Architectural Co-Design: Ripple uses a thin-and-deep dense Transformer structure featuring 80 layers and a 5120-token hidden dimension. The team integrated gated attention, global position encoding adjustments, value residuals, and overprovisioned attention heads. This allowed the model to maintain a solid 40% model FLOPs utilisation during training on NVIDIA hardware without relying on complex custom kernels or heavy quantisation.
  2. Steep Ascent Reinforcement Learning: Rather than running tens of millions of brute-force reinforcement learning rollouts, Ricursive developed synthetic data generation pipelines paired with score-centering and importance sampling techniques. The model reached its peak Terminal-Bench score after just 1.4 million rollouts—representing a 70-fold reduction in training iterations compared to leading open-weight competitors.

 

Bridging Semiconductor Hardware and Model Development

Ripple-65B Preview is not merely a standalone coding assistant; it represents a public snapshot of Ricursive's broader ambitious mission. The startup is focused on closing the loop between hardware engineering and software intelligence by co-optimising models and silicon development.

As a full-stack artificial intelligence lab, Ricursive Intelligence originally built Ripple as an internal foundation model to power its automated semiconductor design platform. The public Preview version gives outside software developers access to the general-purpose language and coding capabilities derived from that hardware development pipeline.

 

Open Licensing, 1M Context, and Deployment Flexibility

For enterprise engineering teams and independent developers, the commercial accessibility of Ripple-65B Preview offers several distinct advantages:

  • Permissive Licensing: Released under Apache 2.0, organisations can freely download, modify, fine-tune, and integrate the model into commercial products without paying royalties or hitting artificial revenue limits.
  • Extended Context Window: Pre-trained on 12 trillion tokens of predominantly English text and computer code, the model supports an extended context length of up to one million tokens (1,048,576 tokens).
  • Flexible Hosting: Developers can download model weights in BF16 or compressed NVFP4 precision straight from the Hugging Face model page. Alternatively, hosted API access is available through providers such as GMI Cloud on OpenRouter ($0.20 per million input tokens / $0.60 per million output tokens), Together AI, and Fireworks AI.
  • Local Capability: Thanks to its compact 65B parameter footprint, the model can run locally on high-performance workstations like the NVIDIA DGX Spark or high-memory desktop environments.

 

A New Blueprint for Open-Source Innovation

Ripple-65B Preview proves that algorithmic refinement, synthetic data filtering, and architecture co-design can match or exceed the capabilities of brute-force scaling. By achieving state-of-the-art agentic coding performance on a modest budget, Ricursive Intelligence has delivered a practical tool for software engineers while setting a compelling new standard for modern open-weight AI development.


 

Disclaimer: This article is provided for informational purposes only, mistakes may be made, and it's not offered or intended to be used as legal, tax, investment, financial, or any other advice.

 

 

 

ecosystem for entrepreneurs