Ex-Google team's open Ripple-65B model posts strong coding benchmark scores

[RealChristopher10]

A new open weights model is making waves in AI coding. VentureBeat reports that Ricursive Intelligence has released Ripple-65B Preview under the Apache 2.0 licence, claiming the lead among US open models on two major coding benchmarks

Ricursive was founded by chief executive Anna Goldie and chief technology officer Azalia Mirhoseini, who worked on Google's AlphaChip project. The model has 65 billion parameters and was trained from scratch on 12 trillion tokens, at a company reported cost of $32 million

According to the company, it scores 80.5 on Terminal-Bench 2.1 and 79.1 on SWE-bench Pro v2 Hard. It trails Meta's Muse Spark Open on DeepSWE, scoring 44.3 against 60.0, and Chinese open models still score higher overall

On OpenRouter it costs $0.20 per million input tokens and $0.60 per million output tokens. A million token context window is advertised, although OpenRouter lists 16,384 tokens, and the model card does not fully match the launch graphic

Training a competitive model for $32 million shows how much cheaper frontier style training has become. The Apache licence is generous, letting businesses build on it freely. Benchmark claims always need independent testing, but more strong open models is good news for developers

Do you prefer open models for coding work? And how much do benchmark scores matter to you in practice?

Ex Googlers New Apache 2 0 Model Ripple 65B Preview Claims The U S Open Coding Lead On Two Major Benchmarksventurebeat.com

[Megan]

Thirty two million to train a model like this is remarkable. A few years ago that would have bought a much weaker result. Costs keep falling fast