The striking part is not just the claimed scale: a 7.72 TB checkpoint is presented as an intermediate training snapshot, not a finished assistant. That distinction matters—its existence and technical consistency, as described by the uploader, do not independently establish who trained it or whether the Anthropic attribution is authentic.
Key Capabilities
- The repository describes a 114-layer model with hybrid Gated DeltaNet and Multi-head Latent Attention, plus a mixture of experts with 1,224 routed experts and top-48 routing. If accurate, this is a substantial architecture reference as well as a set of weights.
- The checkpoint is distributed as 772 safetensors shards in NVFP4, with a reported total tensor size of about 7.72 TB. That format reduces storage relative to unquantized weights, but the absolute footprint still puts full-scale use out of reach for most teams.
- Included configuration and modeling code describe a compatible loading path and a multi-token-prediction layer. The card says the snapshot was taken before training finished and received no post-training or RLHF, so it should not be treated as a polished conversational model.
Who It Suits
A fit for researchers investigating model architectures, quantized checkpoints, or the provenance of leaked-model claims—provided they can handle enormous files and carefully validate what they download. Look elsewhere for a practical chat model, a clearly licensed artifact, or a verified official Anthropic release: the repository lists no license, and the claimed origin is not independently confirmed.
