Ricursive Intelligence发布了专为优化其AI芯片设计平台而训练的前沿开源权重语言模型Ripple-65B Preview 1。该模型在TerminalBench 2.1基准测试中取得80.52的得分,创下美国开源权重模型在智能体编程方面的新纪录 1。这一成绩超越了Reflection Beam(80.1分)和Meta Muse Spark Open(76.4分)等参数规模更大的模型 1。
Ripple-65B Preview的总训练成本为3200万美元,约为前SOTA模型成本的十分之一 1。在架构上,该模型采用密集的Transformer架构而非混合专家系统(MoE),具体配置为d_model=5120且包含80层 1。为实现高计算效率,研发团队采用了精心设计的强化学习配方,仅进行了1.4百万次rollouts,这一数量比Reflection Beam少约70倍 1。此外,该模型基于12万亿tokens的数据进行预训练,支持最高1M的上下文长度 1。目前,开发者可以通过Hugging Face、Together AI、GMI Cloud、OpenRouter以及Fireworks等多个主流AI平台访问并使用该模型 1。
Ricursive Intelligence has released Ripple-65B Preview, an open-weight large language model trained specifically to optimize its AI chip design platform 1. The model achieved a score of 80.52 on the TerminalBench 2.1 benchmark, establishing a new state-of-the-art record for US open-weight models in agentic coding tasks 1. This performance surpasses larger parameter models such as Reflection Beam, which scored 80.1, and Meta Muse Spark Open, which recorded 76.4 1.
A key highlight of the release is its significant computational efficiency and cost-effectiveness. The training cost for Ripple-65B Preview was $32 million, approximately one-tenth of the expense required for previous state-of-the-art models 1. The model utilizes a dense Transformer architecture rather than a Mixture of Experts (MoE) structure, featuring a d_model size of 5120 across 80 layers 1. During development, it was pre-trained on 12 trillion tokens and supports a maximum context length of 1 million 1. Furthermore, the reinforcement learning process employed 1.4 million rollouts, which is roughly 70 times fewer than those used by Reflection Beam 1.
Ripple-65B Preview is currently available through multiple major AI platforms, including Hugging Face, Together AI, GMI Cloud, OpenRouter, and Fireworks 1.
评论
还没有评论,欢迎留下第一条。