2 min · Hacker News
Alleged Anthropic Numbat Checkpoint Details 14T Hybrid GDN and MLA Architecture
A 7.72 TB checkpoint codenamed numbat appeared on Hugging Face with complete PyTorch modeling code and 772 safetensors shards. While the developer community remains deeply skeptical about its authenticity, the included implementation offers a masterclass in hybrid Gated DeltaNet and Multi-head Latent Attention design.
Why it matters. You can inspect the PyTorch reference code to study state-of-the-art hybrid attention layers and NVFP4 dequantization patterns used in frontier model architectures.