Good morning — it's Tuesday, November 3, 2026. The 7.7-terabyte leak that was random numbers, proved fake by a ninety-kilobyte check. From the bench, SenseNova-Loop-9B-Preview runs its layers twice. In the drops: Heretic 2.0, FineDNA's open training data, and an inventory app for the spares shelf. Don't keep us a secret: forward this email to a fellow home labber, or share the web version.

Share this issue:  X   WhatsApp   Reddit   Facebook   LinkedIn   Text   Email

The 7.7-terabyte leak that was random numbers

On Sunday, a Hugging Face repository called burner22/numbat spread across X as a leaked checkpoint from Anthropic's next Mythos-class run. The card claimed about 14 trillion parameters, 114 layers and 1,224 experts across 772 shards totalling 7.72 TB. Created on October 23, it showed 5,846 downloads by Sunday evening. Anthropic has not commented.

The weights were never trained. In the repository's own discussion, a user showed that 43 complete tensors — 730,368 values — regenerate bit for bit from their names: hash the name with SHA-256, seed NumPy's PCG64 generator with the first eight bytes, draw, scale, round. CellCog reran the recipe (independent): nine of nine tensors matched exactly, and a wrong seed matched none. Another auditor extended it to the 4-bit matrices — signs matched on all 162,246,496 values checked — and a Monday rerun (independent) matched 369,408 of 369,408 values. The check downloads about ninety kilobytes.

For your rack, the lesson is the order of operations. Anyone can upload terabytes of plausible weights with a confident card — this one also ships modeling code that asks to be trusted. Read the header first, test a few small tensors, and let provenance do the work file size cannot. Ninety kilobytes of checking beats 7.7 terabytes of downloading.

SenseNova-Loop-9B-Preview — the 9B that runs its own layers twice

SenseNova-Loop-9B-Preview, released by SenseNova on October 30 and trending on Hugging Face this week, is a multimodal model — text and images in, text out — built for long-context document work, coding, and tool use. Its trick is a loop: a 24-layer block of its 32 distinct layers runs twice with shared weights, so each pass executes 56 layers while storing 32 layers' worth. That is how the card (vendor) gets about 14.6B inference parameters out of about 9B stored. Downloads are still around fifteen hundred, and the benchmark scores exist only as a chart on the card (vendor), technical report still to come — no text scores to line up yet, and nothing independent has replicated them.

Setup notes: a custom architecture, so it loads with trust_remote_code under Transformers; serving today means SGLang or vLLM builds taken from the projects' pull requests, and the vendor examples use four GPUs. No Ollama listing on the card. Difficulty: fiddly.

Heretic 2.0 breaks its own config files. Heretic, the training-free tool many local runners use to edit a model's refusal behaviour, shipped version 2.0.0 on November 1 (attributed). The new default method is Arbitrary-Rank Ablation, and the README is blunt: configuration files made for 1.x are not compatible with 2.x, so read the migration guide before you upgrade a working setup.

FineDNA publishes the whole recipe, not just the weights. Hugging Face Bio has released FineDNA, Carbon-X's training data, under Apache 2.0: 514.30B mid-training tokens on the card (the run read 302B), 203,923 supervised conversations, and 98,433 reinforcement prompts whose answers a program can check. The configs total about 950 GB — stream them, says the card.

GitHub find: Kept inventories the spares shelf. Kept is a self-hosted inventory of everything you own and where it is — rooms, shelves, drawers and boxes, each thing with its photos, receipts and warranties, and printed QR labels that open a box's contents. Docker Compose, no telemetry, 472 stars, AGPL-3.0, currently v1.14.3. Point it at the drive-and-cable boxes.

Monday, 11 p.m. Your NAS emails you: the ZFS pool is DEGRADED and one drive has dropped out of it entirely. Everything still reads and writes, a spare drive sits on the shelf beside the rack, and the family photo archive lives on that pool.

Your call:
A. Reseat the dropped drive and force it back into the pool tonight
B. Check the pool status and logs, swap in the shelf spare, and resilver
C. Shut everything down and restore last night's backup to a single drive

100 Hero Points for the right call. Your score is on the outcome page, ready to share.

Reading this because a fellow home labber forwarded it? Get your own copy every morning — subscribe free, and pass it on — that is how this grows.