跳到正文
HNHacker News·
暂不在当前实时榜单

Alleged Mythos Class Anthropic Checkpoint Leak

AI 摘要

An alleged leak of an Anthropic Mythos-class checkpoint, internally codenamed numbat, has surfaced. This checkpoint, from a training run referenced in the Claude Code source leak (Mar 2026), is described as a ~14T parameter, 114-layer model utilizing hybrid GDN/MLA and MoE with 1,224 experts. The leak includes configuration files like configuration_numbat.py, a reference implementation modeling_numbat.py, a tokenizer.json, and 772 model.safetensors shards, along with checkpoint metadata and training logs.

时间与来源

时间显示为 UTC

显示时区:UTC

本地时区尚不可用,暂时显示 UTC。

收录当时偏移:UTC+02026年11月1日 17:00 UTC

收录
2026年11月1日 17:00
来源类型
官方发布

本站未收录正文。

前往源站阅读 →
来源·Hacker News·huggingface.co