AI初创公司Recursive正式推出其首个编码模型Flint，该模型在该公司自研的Eureka Machine AI系统的辅助下完成开发 1。这一项目仅耗时四个月即告完成 1，体现了AI自我改进理念的早期实践。
Flint采用专家混合架构，拥有400亿参数但每个任务仅动用其中的25亿参数 1。该模型已作为免费开源项目发布在Hugging Face平台，采用Apache 2.0许可证 1。在性能表现上，Flint在MacBook Pro M4 Pro芯片上的生成速度达约131 tokens/秒，约为对标模型的两倍 1。在SWE-Bench Verified基准测试中，Flint超越了其他参数规模在50亿以下的模型 1。
Recursive在开发过程中对模型进行了低位训练优化，成功将Flint缩小71% 1，几乎未造成性能损失。由公司CEO Richard Socher领导的团队推进此项目，Eureka Machine参与了开发的几乎每个步骤 1。此前，Recursive在5月以45亿美元估值融资6.5亿美元 1，并与亚马逊云服务达成了金额为4.1亿美元的计算交易 1。
Recursive, an AI startup that raised $650 million at a $4.5 billion valuation in May, has unveiled Flint, its first coding model developed primarily with the assistance of its own Eureka Machine AI system 1. The open-source model was completed in just four months under the leadership of CEO Richard Socher and his team 1.
Eureka Machine participated in nearly every step of Flint's development, demonstrating an early practical application of AI self-improvement principles 1. The resulting model features a mixture-of-experts architecture with 40 billion total parameters, though it uses only 2.5 billion parameters per task 1. On a MacBook Pro with an M4 Pro chip, Flint achieves a generation speed of approximately 131 tokens per second, roughly double that of comparable models 1. The model outperforms other models with fewer than 5 billion parameters on the SWE-Bench Verified coding benchmark 1.
Flint is being released free of charge under an Apache 2.0 open-source license on the Hugging Face platform 1. Through low-rank adaptation, the model has been compressed by 71 percent with minimal quality loss 1. Recursive has also entered into a $410 million computing arrangement with Amazon Web Services to support the project 1.
评论
还没有评论，欢迎留下第一条。