What happened
Ricursive Intelligence, a lab that develops AI across hardware, software and models together, has released Ripple-65B Preview, an open-weight language model it trained from scratch, Ricursive reported. The company describes it as a general-purpose version of a proprietary model that runs its AI chip design platform. Pretraining used 12 trillion tokens, and the model handles context of up to 1M tokens.
On TerminalBench 2.1 the model scored 80.52, which the company says is the best result for any U.S. open-weight model. Ricursive compared it with larger U.S. open-weight systems: Reflection Beam 501B scored 80.1, Meta Muse Spark Open 816B reached 76.4, while NVIDIA Nemotron 3 Ultra 550B posted 56.4 and Thinking Machines Inkling 975B got 63.8. The model can be accessed through OpenRouter, Hugging Face, GMI Cloud and Together AI, with Fireworks due later the same day.
Ricursive put total development cost at $32M, covering all R&D from scratch, and estimated this at about one-tenth of the cost of the previous open-weight leader. It chose a dense Transformer over Mixture-of-Experts and reported 40% MFU during training. The company said it reached its score with 1.4M reinforcement learning rollouts, roughly 70x fewer than Reflection Beam, and built its own synthetic data instead of buying post-training data.
Context
The company set its budget against reported industry costs. OpenAI CEO Sam Altman said in 2023 that training GPT-4 cost more than $100 million, and Dario Amodei later said frontier models could cost up to $1 billion, projecting $10-100B for pretraining by 2026 to 2027. Ricursive also cited reports that Reflection agreed to spend about $150M a month on compute using SpaceX Colossus.
The briefing
The week in tech, in one email.
The dispatches worth your time, every week. Unsubscribe anytime.