据Hacker News社区讨论显示,Anthropic公司代号为“numbat”的Mythos级大语言模型检查点疑似被泄露 12。这一事件被视为继2026年Claude Code源码泄露后,再次证实Anthropic下一代Mythos级模型存在的证据 2。泄露文件包含约14万亿参数、共114层的架构细节,以及总大小约为7.72 TB的权重数据 2。
具体而言,该模型采用混合GDN/MLA架构和MoE(Mixture of Experts)结构,拥有1,224个专家网络,激活方式为top-48加1个共享专家 2。数据格式为NVFP4量化版本,分为772个shard 2。这是一个预训练中途的检查点,步数为Step 1,830,000(总目标约196万步),未包含后训练或RLHF数据 2。上传者还提供了configuration_numbat.py和modeling_numbat.py参考实现代码,并验证了张量在safetensors文件中的加载一致性 2。
有用户评论指出,下载者可能通过分析.safetensor文件的组织结构来推断模型架构以实现推理 1。目前信息主要来源于Hacker News用户的发布与讨论,尚无官方确认 12。
An anonymous user on Hacker News posted a large language model checkpoint claimed to be from Anthropic's internal "numbat" project, which is reportedly part of the company's Mythos-class training run 2. The alleged leak includes approximately 14 trillion parameters across 114 layers, utilizing a hybrid GDN/MLA architecture and a Mixture-of-Experts (MoE) system with 1,224 experts 2. The data footprint totals roughly 7.72 TB in NVFP4 format, distributed across 772 shards 2. According to the source, this checkpoint represents an intermediate stage of pretraining at Step 1,830,000 out of a projected total of approximately 1.96 million steps, and it does not contain post-training or RLHF data 2.
To substantiate the claim, the poster provided reference implementation files named configuration_numbat.py and modeling_numbat.py, which were used to verify the consistency of tensor loading within the safetensors format 2. While related discussions on Hacker News focused on how downloaders might infer the model's architecture by analyzing these .safetensor files, the thread highlights the alleged presence of the leak on Hugging Face 12. This event is described as following a previous source code leak of Claude Code in 2026 and is viewed by some as additional evidence confirming the development of Anthropic’s next-generation Mythos-class models 2.
评论
还没有评论,欢迎留下第一条。