Zincrom sparse-MoE manifests and architecture metadata

This is a local draft for a possible Hugging Face metadata repository. It contains no model weights and is not yet approved for publication.

No model-weight binaries are stored or redistributed in this bundle. Users must obtain base-model weights from the original upstream publishers under the applicable licenses. License metadata for this future repository must remain unset until the contents and upstream obligations receive human review.

Evaluated architectures

The empirical evaluation closed on 2026-07-25 and ran on one Oracle OCI A1 host with four Arm Neoverse-N1 cores and 24 GiB RAM.

OLMoE-1B-7B-Instruct

• Upstream repository: allenai/OLMoE-1B-7B-0924-Instruct
• Tested quantization label: Q4_K_M GGUF
• Architecture: 64 total experts, top-8 routing, 16 MoE layers
• Tested routed-cache policy: per-layer LRU capacity 8 for the exact evaluated model/build
• Measured coordination result: 74 µs p99 BEAM dispatch versus 2.206 ms p99 selected-expert compute

Granite-3.1-3B-A800M-Instruct

• Upstream repository: ibm-granite/granite-3.1-3b-a800m-instruct
• Tested quantization label: Q4_K_M GGUF
• Architecture: 40 experts, top-8 routing, 32 layers
• Tested routed-cache policy: per-layer LRU capacity 4 for the exact evaluated model/build

Qwen1.5-MoE-A2.7B-Chat

• Upstream repository: Qwen/Qwen1.5-MoE-A2.7B-Chat
• Tested quantization label: Q4_K_M GGUF
• Architecture: 64 routed experts plus 4 shared experts, top-4 routing, 24 layers
• Tested routed-cache policy: per-layer LRU capacity 4 for the exact evaluated model/build

These policies are bounded measurements, not universal defaults. Unknown model/build combinations require a conservative fallback and new validation.

Provenance status

The accompanying weight_manifest.json deliberately records unresolved source hashes as null. The original upstream GGUF files are no longer present on the evaluation host, so their exact SHA-256 hashes cannot be reconstructed from the current repository. No derivative shard is included.

Before publication, a human reviewer must verify:

1. exact upstream model repositories, filenames, revisions, and licenses;
2. SHA-256 hashes of newly obtained source files;
3. licenses and notices for every script and metadata file in the final bundle; and
4. consistency with the approved technical report and reserved DOI, if any.

Scope boundaries

• FFT spectral routing is inconclusive.
• The cognitive-core proposal is untested.
• Runtime 2.0 is later work in progress and is not part of the completed empirical evaluation.
• Results apply only to the tested single-host envelope and recorded configurations.

References

• Technical-report draft: docs/publication/technical_report.md
• Evidence authority: docs/CURRENT_EVIDENCE.md
• Claims summary: docs/ZINCROM_CLAIMS_TEST_RESULTS_SUMMARY.md
• Development repository: https://github.com/mailtotanvir/zincrom