RumaStack/GLM-5.3-Flash-NVFP4-AXL-ABLIT

GLM-5.3-Flash-AXL-ABLIT

Release upload in progress: do not download or serve this repository until the release status and pinned revision are announced. The source checkpoint's 31-file set and all 18 shard SHA-256 digests were verified against its original manifest before transfer. The complete remote distribution has not yet been verified; no full consumer download was tested. See Japanese.

What this artifact is

An unofficial static derivative of the public AXL checkpoint with selected donor self_attn.o_proj payloads. It is not a full copy of the Dealign model and is not the separate runtime BF16 overlay. Z.AI, NVIDIA, Bizuayeu and Dealign are upstream sources, not claimed endorsers of this derivative.

The published qualification report identifies the following input revisions. The original local manifest matches the AXL, NVIDIA and donor pins. The Z.AI revision identifies a license notice, not a direct model input.

Role	Model repository	Revision
AXL base	Bizuayeu/GLM-5.3-Flash-NVFP4-attn-lmhead-W4A16	8da41003c93593de40bf4a48e27f8d3f41973ad7
NVIDIA parity source and AXL ancestor	nvidia/GLM-5.3-Flash-NVFP4	423acf37583782c51c142d145aef733d72943d93
Donor	dealignai/GLM-5.3-Flash-UNCENSORED-NVFP4	745aac2ff0f10acf961f396df3f9418598aa7327
Original-model license notice	zai-org/GLM-5.3-Flash	eb9eb208eb0d988989d07a6a12d0fdeb5f52574a

The original-model notice revision is a notice provenance record; it is not an assertion that this derivative directly downloaded or trained on that full checkpoint.

Changes and invariants

The documented static route requantizes main layers 15–43 o_proj from donor BF16 to AXL-compatible W4A16 NVFP4: 29 weight/scale/global-scale triplets, or 87 processed keys. Native MTP layer 45 contributes one BF16 o_proj.weight key. In the reported artifact, that MTP payload is byte-identical to the AXL base. Therefore 88 processed keys does not mean 88 byte-changed tensors, and it does not mean that MTP was newly quantized.

The report states that other tensors, index mapping, configuration, tokenizer, chat template and inherited auxiliary files remain unchanged from the pinned AXL base. No new training-data stage is claimed for this static conversion; that is not a claim that the upstream models required no data or have a fully audited training chain.

License

The weight distribution is MIT, retaining the original text and copyright notice in LICENSE. The source notices and their exact revision identifiers are in UPSTREAM_NOTICES.md. No research-only or non-commercial license restriction is added by this derivative. Applicable law and third-party rights are not waived.

The setup and conversion source repository retains Apache-2.0 and its component-specific notices. Its code license is not a blanket license for the weights, container dependencies or third-party applications. No guarantee of non-infringement, safety, correctness or fitness is made.

Runtime and package boundary

Use the reviewed fork's static derived-checkpoint route and its required runtime overlays. The distribution places the original, unmodified checkpoint under checkpoint/ and keeps this card and licensing notices outside that directory. Point the runtime at the child checkpoint directory. Do not add new files to its manifest-protected file set or edit its inherited README/configuration.

This layout is not a tested generic Hub-root from_pretrained or stock vllm serve configuration. Check release status before using the download and verification instructions, the outer file-hash inventory and the read-only verification script. The runtime path is the checkpoint/ child directory. Do not change production paths or restart services for this publication.

Evaluation and limitations

The published report records an integrity check and experimental two-node TP=2/MTP inference on 2026-09-27. This review verified all shard digests and inherited auxiliary bytes against the original manifest, but did not rerun inference. It is not a production-acceptance record and does not establish answer quality, complete refusal removal, long-duration stability, speculative-token acceptance rate, or superiority to standard AXL. Neither upstream model-card benchmarks nor donor results are asserted as results for this derivative.

Converter provenance correction: the earlier qualification report's 3e3cd0e37706a63e5959baf0f38d7b1268b83d6e42ce48f5b7f66c98c54669ef identifies the source file at its previously cited baseline 2d6472588742f75238e412f2a633d32447790179, not the original manifest's module. Both manifest converter fields and the original reproduction report identify cc8e05b22d38fc6b27c31521105972a3d3b5d1d6d44e866e6abfc44fcf95c32b, matching the public source file at the immediate child a9ee7b4ae99e3627665cc730d7c13e94bee0a53f and current checkout. The intervening diff caches headers during verification, without changing target keys or quantization math. The exact generation-time execution HEAD is not attested; matching file bytes do not establish the repository-wide execution revision. See the historical qualification report for the original attribution, and PROVENANCE.public.json for the corrected local evidence. Neither manifest nor converter was rewritten.

PROVENANCE.public.json separates source-reported expectations from locally observed metadata; it is not the generated runtime manifest. The original manifest and the checkpoint belong exclusively inside checkpoint/, not in this outer documentation set. A privacy screen found no private-data indicators in the original manifest or checkpoint text candidates, but it is not a guarantee against every possible disclosure.