AI news ·
Ripple-65B Preview open model tops US coding benchmark at $32 million training cost
Ripple-65B Preview hit 80.52 on TerminalBench 2.1, the highest among open US models, after training on 12 trillion tokens at a total cost of $32 million.
Ricursive Intelligence, an AI chip design firm, has released Ripple-65B Preview, an open language model that scored 80.52 on the TerminalBench 2.1 coding benchmark - the highest result among open US models. The model was pretrained on 12 trillion tokens and is available under the Apache 2.0 license, with total development and training costs reported at $32 million.
Architecture and training approach
Ripple-65B Preview uses a dense Transformer architecture with 65 billion parameters. This differs from the mixture-of-experts designs common in many recent models. Ricursive optimized the model for Nvidia hardware, which the company said should make it easier to deploy on users' own systems compared to some competing models.
The model partly uses Terminus 2 and partly OpenCode as its evaluation harness. Ricursive said the $32 million figure covers all research and development expenses, from initial work through final training.
Benchmark performance and comparisons
On TerminalBench 2.1, Ripple-65B Preview reached 80.52. The newly released Inkling 2 Small outperforms Ripple across several coding benchmarks, but Ricursive emphasized that its model's architecture and optimization choices prioritize practical deployment.
On SWE-Bench Pro v2.0 Hard, Ripple scored 79.1% against Inkling 2 Small's 86.3%. The gap widened on DeepSWE v1.1, where Ripple reached 44.3% and Inkling hit 72.1%. On MCP Atlas, Ripple achieved 71.67%, while Inkling 2 Small reported scores of 85.4 and 85.0 on two configurations.
Availability and licensing
Ripple-65B Preview is available on Hugging Face under the fully permissive Apache 2.0 license. The open licensing means teams can use, modify, and distribute the model without commercial restrictions, which matters for organizations building agentic coding workflows into their own products.
For developers working on AI code generation, models with strong benchmark results and permissive licenses open up options for fine-tuning and self-hosted deployment. The model's dense architecture also simplifies the hardware requirements compared to mixture-of-experts alternatives that need more complex inference setups.
Why this matters for product development and IT teams
Ripple-65B Preview gives teams a competitive open model for agentic coding tasks without the licensing constraints that limit how commercial models can be integrated into products. The $32 million training cost also signals that capable coding models are becoming achievable at budgets that don't require hyperscaler-level investment. For IT and development leads evaluating build-versus-buy decisions for coding assistants, an Apache 2.0 model that scores above 80 on TerminalBench changes the calculus on what's feasible to run in-house.