Mamba research: frozen project checkpoints

This is a public artifact archive, not a ready-to-use pretrained language model. The Apache-2.0 grant covers the project-produced checkpoint weights and newly authored archive metadata only. Existing source and third-party material retain their original license/copyright status; see SOURCE_AND_LICENSE_SCOPE.md.

Delivery status: 24 of 25 selected weights published

As of 2026-10-07, 24 terminal checkpoint files are actually published. All 24 passed immutable remote hash/size verification and full-byte SHA256 readback. The other selected checkpoint is not available from this repository:

checkpoints/official_zoology_mqar_data/v100_full_iclr24_L64_P4_retimed_s11_lr4e4_paced_v1_background/checkpoint_final.pt

Its size is 13,617,805 bytes; expected SHA256 is ab4a0b2eac71dd27744d0930ee1a7b167fb0662263d04a0476cba377983d3edf. Bounded local and split local/cloud storage-upload attempts failed with timeouts, connection resets, a proxy error, or HTTP 502. This is an upload transport failure, not a training failure or a model-quality exclusion. The completed original weight remains retained locally, as do all 25 original terminal weights. No claim of complete 25/25 delivery is made.

See ARCHIVE_DELIVERY_STATUS.json for the exact 24 published identities, immutable hash-verification revision, per-weight full-byte readback revisions, and the missing item. ARCHIVE_MANIFEST.json remains the unchanged historical 25-item intended selection, not proof that all listed files were uploaded. It also records the original README hash; that original text is preserved in archive_status/README_frozen_pre_completion_20261007.md. This README was subsequently corrected to report actual delivery.

What is included

  • 19 published completed, full-budget single-cell checkpoints on pinned official Zoology MQAR data: project endpoint (7), project retimed (5), attention comparator (6), and a partial FP32 SISO reference (1). The intended selection contained one additional retimed checkpoint, identified as missing above.
  • Five historical endpoint mechanism checkpoints, training seeds 11--15, from a bespoke supplementary key/value diagnostic.
  • Terminal checkpoints, exact-source minimal subsets, effective configs, provenance, completion/summary evidence, original and published-file hashes.

The terminal selection is archival, not selection of favorable validation checkpoints. Completed runs with poor outcomes are retained. These are only selected historical runs, not all preregistered experiments or the entire evaluation record. Initial/epoch/recovery weights and raw arrays/datasets are outside this model repository; their separate retention/archival decisions are not described by this manifest. Redundant upload-staging copies may be removed after remote hash/full-byte verification while the original terminal weights remain retained. No training or model execution was performed while preparing this archive.

Architecture and research limits

Project endpoint/retimed models are not complete official Mamba-3. The SISO reference excludes MIMO and other features explicitly documented in its captured PROVENANCE.json. Full-budget MQAR cells are not full reproduction of the original Mamba-3 paper or proof of generalization, novelty or broad superiority. The bespoke mechanism checkpoints are supplementary diagnostics, not standard benchmark substitutes. These historical checkpoint families do not themselves close the original-paper reproduction gate; their documented gradient/cache/ numerical and full training/protocol limitations remain.

The official 1.5B MIMO checkpoint is not mirrored or relabeled. Use the pinned upstream reference in UPSTREAM_REFERENCES.json. That checkpoint is separate from the project models archived here.

File layout and retrieval

Each selected run directory under checkpoints/ has config/provenance and source_snapshot_minimal.zip; 24 also have the completed terminal .pt, while the missing weight's directory contains metadata only. The subset zip is newly derived from the run's archived source and includes only hash-verified selected files. It is not the original full source archive. Its subset manifest records omitted paths; original source ZIPs remain local because earlier whole-vendor license audits found unresolved third-party redistribution issues.

Download a selected file with huggingface_hub.hf_hub_download, pin the archive commit, and compare SHA256 with ARCHIVE_MANIFEST.json before loading. .pt files may include optimizer/scheduler/RNG metadata and are not advertised as exact continuation states. Use a trusted PyTorch environment and torch.load(..., map_location="cpu", weights_only=True) where supported; this archive does not authorize unsafe pickle loading and does not certify the checkpoint-loading/runtime gate. Do not assume Transformers from_pretrained compatibility.

The exact original metadata hashes and checkpoint bytes are retained. Published metadata is a derived JSON rendering with local user/workspace paths redacted. Local interrupted/failed experiments and prior manuscripts were not rewritten.

Attribution

See UPSTREAM_REFERENCES.json, ZOOLOGY_LICENSE.md, NOTICE, and the per-checkpoint source subset manifests. The research source repository is https://github.com/guangxiangdebizi/mamba-research . This archive is not endorsed by the Mamba or Zoology authors. No scientific claim follows from successful hosting or hash verification alone.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support