Milo-Ark is the checksummed model store on miloshare. Preservation is the contract. Offline serve is recorded only when a smoke result sits beside the checksums. Storage rebuild lives in the DS1823xs+ post.

70checksummed trees
27.46 TBchecksummed payload
13 TiBfree of 40 TiB
1smoke test in a manifest
4GB300 serving notes

Now: every tree in the Ark is checksummed and nothing is downloading. Fourteen trees joined the inventory since September 19, and the 3.77 TB audnai/penclaw-GLM-5.3-abliterated tree passed its NAS-side verify (738/738). The biggest new tree is moonshotai/Kimi-K3.1-Pro (1.56 TB, 98/98 LFS hashes checked on the NAS against the Hub). deepseek-ai/DeepSeek-V4.1-Flash-1017 is the first tree whose manifest records a passed smoke test: vLLM on the GB300 Station, tool gauntlet 12/12, GSM8K-200 96.5%, needle-checksum 500/500, NAS replica 85/85. That serve ran from the Station's NVMe copy of the same revision, so it is not yet an offline-from-the-Ark serve.

Closed today: meta-models/Muse-Spark-Open-816B-A42B-NVFP4 (533.0 GB) was hashed on the NAS and all 128 large files match the Hub's sha256. The stalled abenzerps/Qwen-Image-2.1-Uncensored-GGUF copy was finished from the local stage and matches the Hub 17/17. nvidia/DeepSeek-V4.1-Flash-NVFP4 had checksums but no record of a check on the NAS; sha256sum -c on the NAS now passes 83/83.

27.5 TB checksummed across 70 trees Bytes on the NAS by model family, split by who published them. Measured November 13, 2026. GLM 12.08 TB · 21 trees Kimi 7.40 TB · 8 trees DeepSeek 4.76 TB · 14 trees Qwen 1.17 TB · 17 trees MiniMax 0.70 TB · 3 trees MiMo 0.57 TB · 1 tree Other 0.78 TB · 6 trees model lab NVIDIA quant community / derivative
Figure 1. Bytes per model family across the 70 checksummed trees. GLM is the largest family (12.08 TB in 21 trees), mostly community derivatives such as the 3.77 TB penclaw abliteration. "Model lab" means the original publisher (zai-org, deepseek-ai, moonshotai, Qwen, XiaomiMiMo, google, NousResearch).

Checksummed inventory

RepoRevSizenState
abenzerps/Qwen-Image-2.1-Uncensored-GGUF new6b34e59458d3105.2 GB25checksum
audnai/penclaw-GLM-5.3-abliterated new027ae98a6f9d3.77 TB739checksum
avlp12/Kimi-K2.7-Code-Alis-MLX-Dynamic-3.6bpw2407e9b7948a464.9 GB118checksum
Blackfrost-AI/DeepSeek-V4.1-Flash-1017-DERISKED-FP8 new4bb7df57c6bf510.3 GB88checksum
Blackfrost-AI/KIMI-K3-Q2_K-GGUF-ABLITERATED newb233c5abd8be1.01 TB51checksum
Blackfrost-Research/GLM-5.3-F.U-AnthraClaud-Edition-BF16 neweac6aa06d2b11.51 TB169checksum · >HW
Blackfrost-Research/GLM-5.3-Flash-DERISKED-NVFP4ac08b5ee7c8b205.1 GB133checksum
cebeuq/DeepSeek-V4-Flash-0731-abliterated new21bd923c2574168.4 GB89checksum
dealignai/DeepSeek-V4.1-Flash-UNCENSORED-FP8d61c59ea5e51510.3 GB62checksum
dealignai/GLM-5.3-ABLITERATED-NVFP4371bdb985d01464.9 GB294checksum
dealignai/GLM-5.3-Flash-ABLITERATED-FP83eb1880f95b1328.4 GB75checksum
deepseek-ai/DeepSeek-V4-Flash60d8d70770c6159.6 GB73checksum
deepseek-ai/DeepSeek-V4-Flash-07317872f01b1d1f166.9 GB78checksum
deepseek-ai/DeepSeek-V4-Flash-DSpark913f0657a874166.9 GB74checksum
deepseek-ai/DeepSeek-V4-Flash-Vision-Exp6821d6ad3681167.8 GB84checksum · served
deepseek-ai/DeepSeek-V4-Prob5968e9190ef864.7 GB91checksum
deepseek-ai/DeepSeek-V4.1-Flashdf42c109f1de510.3 GB88checksum · served
deepseek-ai/DeepSeek-V4.1-Flash-1017 new1bae0e108899510.3 GB85checksum · smoke passed
drowzeys/keys-DeepSeekV4-Flash-GA-0731-Dspark-Abliterated-32-326c47916f85e5166.9 GB80checksum
drowzeys/keys-DeepSeekV4-Flash-GA-0731-Dspark-Abliterated-Anchored-Tensors newa1e693795173166.9 GB81checksum
drowzeys/keys-DeepSeekV4Flash-Vision-EXP-ablit new48095b3452a1167.8 GB85checksum
drowzeys/keys-GLM-5.3-Flash-NVFP4-ablit-l15-43-mtp-l4580b6d18d77e3197.9 GB34checksum
google/gemma-4-26b-a4b-it20da991ab4af51.6 GB12checksum
incoai/GLM-5.3-DFlash2425aa615ce324.9 GB9checksum
incoai/GLM-5.3-Flash-DFlash27d74cdd881ed2.3 GB8checksum
incoai/GLM-5.3-NVFP454e52520606f464.9 GB101checksum
incoai/Qwen3.8-27B-DFlash2adde41d8fde33.8 GB9checksum
k-l-lambda/kimi-k2.7-code-eagle3-mlae2b24ca469934.1 GB4checksum
LibertAIDAI/GLM-5.3-Flash-NVFP4aa28e1f54130194.7 GB134checksum
lmstudio-community/GLM-4.5-Air-MLX-8bit251aed6f354a113.6 GB32checksum
lmstudio-community/MiniMax-M2.5-GGUF93cbd1a0bf00430.9 GB14checksum
lmstudio-community/MiniMax-M2.5-MLX-4bit9af822e97aaa128.7 GB35checksum
lmstudio-community/Qwen2.5-Coder-32B-Instruct-MLX-4bit96572f69fd8818.4 GB14checksum
lmstudio-community/Qwen2.5-Coder-32B-Instruct-MLX-8bit7bd37ed2a7f034.8 GB17checksum
lmstudio-community/Qwen3.5-397B-A17B-MLX-4bit76a66de5151f223.9 GB58checksum
local-inference-lab/Qwen3.8-Flash-Next-NVFP4-4p89ee0cea634a37105.9 GB55checksum
malaiwah/GLM-5.3-Flash-TR3-6bpw1bdd7d3d9aea253.7 GB268checksum
meta-models/Muse-Spark-Open-816B-A42B-NVFP4 new4239e4db1937533.0 GB146checksum
mlx-community/gemma-3-12b-it-qat-4bit66fc51ef25778.1 GB15checksum
mlx-community/magnum-v4-72b-4bit8c5ef3197ee040.9 GB18checksum
moonshotai/Kimi-K2.7-Code74797c9c6237595.2 GB87checksum
moonshotai/Kimi-K3a590ce090cb01.56 TB122checksum · >HW
moonshotai/Kimi-K3.1-Pro new6d0a020295671.56 TB116checksum
NousResearch/Hermes-4-70Bd5dec2bd6b39141.1 GB39checksum
nvidia/DeepSeek-V4.1-Flash-NVFP43431dde3247c527.3 GB83checksum
nvidia/GLM-5.2-NVFP4aec724e8c7b8464.9 GB57checksum
nvidia/GLM-5.3-Flash-NVFP409b04e5e74bc204.5 GB45checksum · served
nvidia/Kimi-K2.7-Code-NVFP49f28d60908c9595.2 GB87checksum
nvidia/Kimi-K3-NVFP4b2428a0b83a81.61 TB120checksum · >HW
nvidia/MiniMax-M2.7-NVFP4e79701cb1f9d139.9 GB30checksum
nvidia/parakeet-tdt-0.6b-v21b149a3589352.5 GB3checksum
nvidia/Qwen3.8-Flash-Next-NVFP4fc694b54fb01132.7 GB25checksum · served
Qwen/Qwen-Image-2.1 newb3179ad355be33.1 GB29checksum
Qwen/Qwen3-Coder-30B-A3B-Instructb2cff646eb4b61.1 GB28checksum
Qwen/Qwen3-Coder-Nexta7fbcb5c0e12159.4 GB52checksum
Qwen/Qwen3-Embedding-8B1d8ad4ca9b3d15.2 GB17checksum
Qwen/Qwen3-Reranker-4B22e683669bc08.1 GB17checksum
Qwen/Qwen3-VL-30B-A3B-Instruct9c4b90e1e4ba62.2 GB25checksum
Qwen/Qwen3.6-27B6a9e13bd6fc855.6 GB29checksum
Qwen/Qwen3.6-35B-A3B995ad96eacd971.9 GB40checksum
Qwen/Qwen3.8-27B1d4bf0f2ff6055.6 GB32checksum
RadixArk/GLM-5.3-NVFP4 new6e389189d564464.9 GB58checksum
RedHatAI/GLM-5.3-Flash-NVFP4240131d6a447197.9 GB23checksum
unsloth/Qwen3.5-35B-A3B-GGUFbc014a17be4322.0 GB5checksum
XiaomiMiMo/MiMo-V2.6-Pro-RL new54b10491b181573.5 GB158checksum
zai-org/GLM-5.3 newba32f8c19c8731.6 MB15checksum · delta
zai-org/GLM-5.3e0b07fd2751b755.7 GB158checksum
zai-org/GLM-5.3-BF16304b8051cfb21.51 TB295checksum · >HW
zai-org/GLM-5.3-Flash04c4e9e95c5d328.4 GB76checksum
zai-org/GLM-5.3-Flash-BF16f12e0fe1f6b2642.7 GB134checksum

new marks trees added or verified since September 19. served means a GB300 serving note in that tree's manifest (Vision-Exp, V4.1-Flash, nvidia GLM-5.3-Flash-NVFP4, nvidia Qwen3.8-Flash-Next-NVFP4). smoke passed means the manifest records a passed smoke test. >HW means runtime proof is N/A because the snapshot exceeds owned hardware. delta is the zai-org/GLM-5.3 chat-template revision: its 141 weight files are byte-identical to e0b07fd2751b, so only the changed small files are stored. Sizes are bytes in original/; n is SHA256SUMS lines. An M3 Ultra parking lot under M3UltraModels/ is not in this table.

Quarantine

RepoWhy
inferencerlabs/Kimi-K2.6-MLX-3.6bitsource gone; shards 44/48 and 47/48 missing
DevQuasar/moonshotai.Kimi-K2.5-GGUFonly shard 40/42; size mismatch
lmstudio-community/gemma-4-26B-A4B-it-GGUFthree GGUF sizes ≠ pinned revision

Done when: pinned files + SHA256 + provenance, then a local-files-only serve and smoke test stored beside the checksums. Checksum ≠ runtime proof. Serving notes ≠ archive smoke files. A smoke test on a staged copy ≠ a serve from the Ark itself. SHA256SUMS written ≠ sha256sum -c passed.