Ricursive Intelligence developed the 80-layer dense Transformer model from scratch for an estimated $32 million, roughly one-tenth the cost of the prior benchmark leader. Pretrained on 12 trillion tokens with support for up to a 1-million-token context length, it achieved an 80.52 score on TerminalBench 2.1. The system attained this benchmark through 1.4 million reinforcement learning rollouts rather than external posttraining data. The weights are currently accessible across platforms like Hugging Face, Together AI, GMI Cloud, and OpenRouter.
Key points
Efficient posttraining techniques and synthetic data pipelines can drastically reduce the…: Efficient posttraining techniques and synthetic data pipelines can drastically reduce the financial capital required to build cutting-edge coding models.
The 65-billion-parameter footprint enables high-performing agentic coding tools to deploy…: The 65-billion-parameter footprint enables high-performing agentic coding tools to deploy on local hardware like NVIDIA DGX Spark devices.
Immediate hosting across platforms including Hugging Face and Together AI gives developer…: Immediate hosting across platforms including Hugging Face and Together AI gives developers broad access to evaluate the 1-million-token context window.
Sources checked by our analysis pipeline: 1
Introducing Ripple-65B Preview,
a Frontier Open-Weight Model
Ricursive Intelligence is a full-stack artificial intelligence lab building self-improving systems by co-optimizing hardware, software, and model.
To power our chip design platform, we trained a language model from scratch that sets a new U.S. open-weight SOTA on agentic coding. We optimized the model architecture for our training and serving platforms, driving substantial efficiency gains.
Today, we release Ripple-65B Preview, a general-purpose variant of the proprietary model that powers our AI chip design platform. The model is pretrained on 12 trillion tokens with support for up to 1M context length. Ripple-65B Preview delivers strong results on agentic coding, achieving U.S. open-weight state-of-the-art (80.52) on TerminalBench 2.1 and surpassing significantly larger US open-weight models such as Reflection Beam 501B (80.1), Meta Muse Spark Open 816B (76.4), Thinking Machines Inkling 975B (63.8), and NVIDIA Nemotron 3 Ultra 550B (56.4).
Ripple’s dense model architecture provides strong reasoning performance with low-latency token generation. The model’s small parameter envelope also enables it to run efficiently on local devices such as NVIDIA DGX Spark, which feature high compute FLOPs…
Developers and investors are adjusting to an altered domestic operating landscape following recent political shifts across the United States. While overall spending across the sector will remain steady, individual projects face potential delays, cost increases, and site relocations. These emerging pressures will fundamentally dictate where and how new facilities are established nationwide.
Internal cost-management measures have been instituted regarding the technical resources allocated to staff members. According to recent reports, restrictions have been implemented targeting the daily financial expenditures permitted for artificial intelligence tools. These adjustments specifically impact the budgetary allowances granted to technical personnel.
Qualcomm has introduced the Cloud AI 100 Ultra, an artificial intelligence accelerator card designed for high-performance workloads. The hardware operates at a 150-watt power level to deliver AI processing capabilities. The accelerator is equipped with 128 gigabytes of memory.