Small open model Ripple-65B beats giants on coding
Ripple-65B Preview is a new open-weight model that claims state-of-the-art performance on TerminalBench 2.1, reportedly beating much larger open models including Meta's 816B Muse Spark and Nvidia's 550B Nemotron 3 Ultra. Ricursive says it trained the model for about $32 million, roughly a tenth of the estimated cost of the previous best open agentic coding model. It is available on Hugging Face, Together AI, GMI Cloud, and OpenRouter.
Why it matters
๐ป Developer ยท A much smaller, cheaper open model matching or beating 500B+ parameter rivals on agentic coding is worth benchmarking against your current stack.
๐ฆ Product ยท Cheaper training costs for frontier-level coding performance could lower the price of AI coding features for smaller companies.
๐จ Design ยท Not directly relevant, but it reflects the broader trend of smaller, more efficient models catching up to giants.
๐ Business ยท If the $32 million training cost claim holds up, it undercuts assumptions that only massive labs can produce frontier coding models.
๐ค Just Curious ยท A relatively small AI model trained on a modest budget is reportedly beating models over ten times its size at writing and fixing code.
Sources: Ripple-65B Preview