Alleged 14T-param Anthropic 'numbat' pretraining checkpoint leaked to Hugging Face
LatentSpacer · reddit · 2026-11-02
A Reddit user claims a checkpoint of Anthropic's next Mythos-class training run was leaked to Hugging Face (unverified).
- Codename numbat: 14T params, 114 layers, hybrid GDN/MLA attention, MoE (1,224 experts, top-48 + 1 shared), NVFP4 quantization
- 772 safetensors shards totaling 7.72TB, plus a reference implementation, 131k-token tokenizer, training log tail and checkpoint metadata
- Claimed to be step 1.83M of a 1.96M-step pretraining run — no post-training/RLHF; the NVFP4 copy allegedly runs evals on 85 Trainium2s
- Cited evidence: numbat appeared in the March 2026 Claude Code source leak; all tensors are internally consistent under NVFP4 round-trip
- Uploader won't disclose the source; authenticity unconfirmed
More from Models
- amasad warns subsidized AI subscriptions will get 'rug pulled' as Grok users hit limits faster — amasad · 2026-11-02
- vLLM serves DeepSeek-V4.1-Flash at 954K tokens/s per GPU on one GB300 NVL72 rack — bookwormengr · 2026-11-02
- OpenAI now caps paid ChatGPT accounts per device and payment card — Alone-Dragonfruit602 · 2026-11-02
- Tracker: Chinese open-weight models ~6-8 months behind frontier, gap widening past a year — scaling01 · 2026-11-02
- Google reportedly testing Gemini 4 Carbon checkpoint on Arena disguised as Argon 4 — testingcatalog · 2026-11-02
- 550M-param AlphaFold3 unbeatable on cost/quality, sparking small-vs-big model debate — nlarusstone · 2026-11-02