Milo-Ark is a local AI continuity archive. Current status: recorded entries are checksum-tracked; runtime proof is still pending. A new GB300 frontier wave is now selected and downloading, but it is not counted as archived until checksum manifests and completion markers exist.
The point is not hoarding weights. The point is being able to answer four operational questions later: what artifact do we have, why do we care about it, can we verify the bytes, and can we actually serve it from our own storage when the upstream copy disappears or changes?
July 9, 2026 update — GB300 frontier wave started. James steered this refresh toward frontier-scale preservation, not small canaries. The active queue is roughly 2.06 TB and targets Blackwell/GB300-practical artifacts where that makes sense. At publish time, this wave was still download-in-progress, so these rows are explicitly not counted as archived yet.
GB300 Frontier Wave — In Progress
| Priority | Repo | Revision | Dry-run size | Status | Why this one |
|---|---|---|---|---|---|
| 1 | nvidia/GLM-5.2-NVFP4 | aec724e8c7b8ee9db3b48c01c320f63f9cdaf8aa | 465.1G | downloading; checksum pending | GLM 5.2 was James's first call; MIT flagship agentic/coding model, with NVIDIA NVFP4 as the practical Blackwell artifact. |
| 2 | deepseek-ai/DeepSeek-V4-Pro | b5968e9190ef611bbf34a7229255be88a0e937c1 | 865.7G | queued; checksum pending | DeepSeek V4 Pro preserves the upstream official 1.6T / 49B-active, 1M-context MIT snapshot rather than only the smaller Flash line. |
| 3 | nvidia/Kimi-K2.7-Code-NVFP4 | 9f28d60908c98034bb288e755566764eac77cc0d | 594.1G | queued; checksum pending | Best third frontier pick: latest Kimi code/agent family, using the Blackwell-practical NVIDIA NVFP4 artifact. |
| 4 | nvidia/MiniMax-M2.7-NVFP4 | e79701cb1f9dce8fe5395b9ed2b20170beebecde | 139.7G | queued; checksum pending | Diversifier hedge: strong agent-team/productivity model, small enough to include beside the three frontier monsters. |
Archive discipline: the July 9 rows graduate from “downloading/queued” to “checksummed” only after SHA256SUMS exists and verifies for that revision. They graduate to “operationally proven” only after an offline serve and smoke test from local files on the target runtime.
The summary cards below describe the prior recorded inventory. The July 9 GB300 wave is tracked separately until verification finishes.
What Is Proven So Far
- Inventory: each entry below exists in the archive tree and has a documented role.
- Integrity: checksum manifests cover the recorded files, so later drift can be detected.
- Coverage: the set spans chat, coding, vision-language, embedding, reranking, speech-to-text, and draft/speculative artifacts.
Model Summary
| Model / artifact | Role | Files | Size | Status |
|---|---|---|---|---|
Qwen/Qwen3.6-35B-A3B | general chat seed | 40 | 67G | checksummed |
Qwen/Qwen3.6-27B | general chat proof candidate | 29 | 52G | checksummed |
Qwen/Qwen3-Coder-30B-A3B-Instruct | coding | 28 | 57G | checksummed |
Qwen/Qwen3-Coder-Next | coding | 52 | 148G | checksummed |
Qwen/Qwen3-VL-30B-A3B-Instruct | vision-language | 25 | 58G | checksummed |
Qwen/Qwen3-Embedding-8B | embeddings | 17 | 14G | checksummed |
Qwen/Qwen3-Reranker-4B | reranking | 17 | 7.5G | checksummed |
google/gemma-4-26b-a4b-it | fallback chat | 12 | 48G | checksummed |
NousResearch/Hermes-4-70B | high-quality fallback | 39 | 131G | checksummed |
nvidia/parakeet-tdt-0.6b-v2 | speech-to-text | 3 | 2.3G | checksummed |
deepseek-ai/DeepSeek-V4-Flash | DeepSeek master | 73 | 159.6G | checksummed |
deepseek-ai/DeepSeek-V4-Flash-DSpark | DSpark companion | 74 | 166.9G | checksummed |
nvidia/GLM-5.2-NVFP4 | GLM 5.2 practical artifact | 57 | 464.9G | checksummed |
moonshotai/Kimi-K2.7-Code | Kimi 2.7 code master | 87 | 595.2G | checksummed |
avlp12/Kimi-K2.7-Code-Alis-MLX-Dynamic-3.6bpw | Kimi MLX practical artifact | 118 | 464.9G | checksummed |
k-l-lambda/kimi-k2.7-code-eagle3-mla | Kimi EAGLE draft | 4 | 4.1G | checksummed |
Definition of Done
A download is not an archive. Milo-Ark marks an entry complete only after a local/offline runtime can serve it from known files and pass a smoke test. The missing artifact is not another table row; it is a reproducible runtime note beside the checksum record.
For each model, the final proof should include the exact archive path, runtime command, host or engine used, smoke-test prompt, response evidence, and any adapter needed to make the model load cleanly. If a serving path requires a converted or staged copy, that staging step needs to be documented too.
Next Step
Pick one proof target, serve it from the archived copy or a documented staging copy, record the runtime command, and save the smoke-test result next to the checksums. After that, repeat the same proof pattern down the table until the archive is not just recoverable, but runnable.