站内导读

中文

定位:numbat 是 Anthropic Mythos 级训练任务在预训练中途导出的 checkpoint,约 14T 参数、114 层,并附带整套参考实现,便于核对张量布局与加载路径。

核心能力

  • 主干为 GDN 与 MLA 混合结构,前馈层用 MoE:每层 1,224 个专家,top-48 路由另加 1 个共享专家。
  • 权重为 NVFP4:E4M3 块缩放配 E2M1 数据,772 个 safetensors 分片,合计约 7.72 TB,单文件不超过 10.6 GB。
  • modeling_numbat.py 定义 NumbatForCausalLM,模块树为每个张量留了位置,可 load_state_dict(strict=True)。
  • forward 跑主干,draft 跑 MTP 层;类继承 PreTrainedModel,可用 trust_remote_code 加载。

README · 详细文档

来源 · 查看原文

numbat

~14T param, 114 layers, hybrid GDN/MLA, MoE (1,224 experts, top-48 + 1 shared), NVFP4. Internal codename numbat, Anthropic's next Mythos-class training run. numbat is the codename referenced in the Claude Code source leak (Mar 2026). This is a checkpoint from that training run, pulled mid-pretraining.

Files:

FileContents
config.jsonmodel configuration
configuration_numbat.pyNumbatConfig
modeling_numbat.pyNumbatForCausalLM reference implementation
tokenizer.jsonbyte-level BPE, 131,072 tokens
model.safetensors.index.jsontensor index across all 772 shards
model-00001-of-00772.safetensorsshard 1 of 772
ckpt_metadata.jsoncheckpoint metadata (ckpt path, eval harness output)
training_log_tail.txtlast ~400 steps of training log

The 772 shards total ~7.72 TB of tensor data. Each shard is safetensors format, under 10.6 GB per file.

configuration_numbat.py and modeling_numbat.py are included: a complete reference implementation of the checkpoint. Its module tree has an exact slot for every tensor in every shard, so the full checkpoint loads with load_state_dict(strict=True) (and, since the classes derive from PreTrainedModel, also via AutoModelForCausalLM.from_pretrained(..., trust_remote_code=True)). NumbatForCausalLM.forward runs the trunk and NumbatForCausalLM.draft runs the MTP layer.

$ python -c "from safetensors import safe_open; f = safe_open('model-00001-of-00772.safetensors', framework='pt'); print(len(f.keys()))"
142

Verification:

$ cat model.safetensors.index.json | python -c "import json,sys; d=json.load(sys.stdin); print(d['metadata']['total_size'])"
7720300817720

Mirrors welcome. Do not DM me asking for the source.

Why believe this?

  • numbat was confirmed as a codename for an unreleased Anthropic model

in the Claude Code source leak (March 2026), found in a comment in prompts.ts: "Remove this section when we launch numbat."

  • The checkpoint's architecture (hybrid GDN/MLA, 1,224 experts, NVFP4)

matches what would be expected of a next-generation Mythos-class training run

  • Every tensor in every shard is internally consistent: E4M3 scale

blocks decode their E2M1 payloads within RTNE bounds, block scales never underflow, and every NVFP4 quad round-trips exactly

  • File sizes, tensor counts, and index entries are all deterministic

and self-consistent across all 772 shards

Loader

The architecture is a hybrid Gated DeltaNet (GDN) + Multi-head Latent Attention (MLA) stack with Mixture-of-Experts feed-forward layers and Multi-Token Prediction. A working single-device reference implementation is included (modeling_numbat.py); the eval runs used the internal Neuron-parallel runtime, but this port consumes the same tensors and round-trips every shard.

FAQ

Q: How did you get this? A: Can't say more than what's already in the repo.

Q: When will more shards be uploaded? A: They're all uploading. Be patient.

Q: Can I use this for inference? A: Yes — modeling_numbat.py implements the full architecture. Load the tensors and run. But this is a pretraining checkpoint at step 1,830,000 of a ~1.96M step run. No post-training, no RLHF. Raw pretraining-completion output. That's expected.

Q: Why is a mid-pretraining checkpoint in NVFP4? A: The eval team quantizes every async checkpoint to NVFP4 so evals run on a slice of Trainium2s instead of the full training cluster — still ~85 of them just to hold these weights. The raw BF16 training weights plus optimizer state never leave the cluster. This is the eval copy.

Q: Is this real? A: Yes.

相关项目 · HuggingFace

查看全部「HuggingFace」项目 →