Editorial illustration for Ripple-65B Preview: fully open Apache 2.0 US model strong at agentic coding
Ripple-65B Preview: fully open Apache 2.0 US model...
Ricursive Intelligence, a company that normally designs AI chips, has put out a language model called Ripple-65B Preview, and it's released under Apache 2.0, meaning anyone can use, modify, or deploy it commercially without restriction. The model was pretrained on 12 trillion tokens and runs on a dense 65-billion-parameter Transformer rather than the mixture-of-experts setup most labs favor these days. Ricursive built it with Nvidia hardware in mind, which matters for anyone trying to actually run the thing outside a hyperscaler's data center.
The headline number is a TerminalBench 2.1 score of 80.52, which Ricursive says is the best mark any open US model has posted on that coding benchmark. Ripple leans on a mix of Terminus 2 and OpenCode as its harness to get there. It's not the top performer across every test, Inkling 2 Small beats it in places, but Ripple's architecture and licensing make it a more practical pick for teams that want to self-host rather than lease compute from someone else. Ricursive also disclosed what it spent getting here, a figure that stands out given how costly frontier model training usually runs.
AI chip design firm Ricursive Intelligence has trained an open language model. Ripple-65B Preview was pretrained on 12 trillion tokens and, according to the company, achieved a score of 80.52 on the TerminalBench 2.1 coding benchmark, the best result among open US models.
Why this matters
A chip design firm training its own frontier-ish model is the real story here, not just the benchmark number. Ricursive Intelligence isn't a typical lab chasing leaderboard glory; it's a hardware company signaling that it wants control over the full stack, from silicon to the models that run on it. For developers and founders, the Apache 2.0 license is the practical headline: no usage restrictions, no API lock-in, full weights to self-host and fine-tune.
That's worth more than a single TerminalBench score, especially since Ripple's 80.52 depends partly on Terminus 2 and partly on OpenCode as harnesses, which makes direct comparisons murkier than the press framing suggests. It also doesn't beat Inkling 2 Small on other benchmarks, so "best open US model" comes with an asterisk about which tasks and which harness. The claimed low training cost deserves scrutiny too, since Ricursive hasn't published full methodology.
Worth watching: whether independent teams can reproduce these agentic coding results outside Ricursive's chosen harness, and whether other chip vendors follow this build-your-own-model playbook.
Related Reading
Research & Benchmarks Microsoft AI Chief: "If It Isn't Safe We Shouldn't Build It
Research & Benchmarks Google's FACTS benchmark shows 70% factuality ceiling across four tests
Research & Benchmarks Nvidia's DLSS 4.5 beta adds 6x Multi Frame Generation for RTX 50 GPUs
From Archives Small team builds model in four months using the Eureka Machine to test ideas
From Archives LangSmith BYOC Expands to AWS, Azure, and Google Cloud
Further Reading
- Ex-Googlers' new Apache 2.0 model Ripple-65B Preview claims the U.S. open coding lead on two major benchmarks - VentureBeat
- Ripple-65B Preview Agent Tests and Results - GMI Cloud
- Introducing Ripple-65B Preview, a Frontier Open-Weight Model - Ricursive Intelligence
- Introducing Ripple-65B Preview, a Frontier Open-Weight Model - AlphaXiv
- Ricursive Says Ripple-65B Preview Scored 80.52 on Coding Benchmark - TokenPost
