GLM-5.3-EXL3-2.75bpw-TP3

GLM-5.3 EXL3 2.75 bpw, TP3 rotating uneven

This is an EXL3/TR3 weight-quantized version of the full GLM-5.3 BF16 checkpoint, not GLM-5.3 Flash. It is not a fine-tune; no additional language-model fine-tuning was performed. It retains all 256 routed experts and official top-8 routing, including the MTP layer. The named bpw is the routed-expert quantization target, not the whole-checkpoint average; passthrough tensors remain at source precision.

Geometry is rotating uneven 768/640/640 with no expert-channel padding. The 768-wide owner rotates across normal MoE layers; MTP layer 78 uses rank 2. K3 uses three-bit routed experts; K2.75 mixes 64 two-bit and 192 three-bit experts per routed layer. These are the original sealed versions, not the rejected scale/re-encoding experiments.

Custom runtime required

This is NOT a stock Transformers, standard vLLM, llama.cpp/GGUF or drop-in ExLlama checkpoint. Use the custom loader, rank-pack preparation, pinned runtime dependencies and safety gates in the reproduction repository. The runtime image/compiled extension are not included here; the source project is an existing-homelab reproduction bundle, not a clean-machine installer.

Lossless transport layout

The sealed checkpoint has more than 19,000 files at its root. To respect Hub folder limits, this repository puts the unchanged original files under checkpoint/part-NNNN/, at most 512 files per part. The original index and manifests are intentionally unchanged and become usable after restoration. TRANSPORT_MANIFEST.json records the exact mapping, byte sizes and SHA256 hashes. The source metadata is also preserved, including its original README and license.

Download the complete repository, then run:

hf download mj-kang/GLM-5.3-EXL3-2.75bpw-TP3 --local-dir ./downloaded-model
python3 ./downloaded-model/restore_checkpoint.py /absolute/path/to/restored-checkpoint --verify

The default restoration links files without copying hundreds of gigabytes. Keep the downloaded snapshot/cache. Use --copy only if you want an independent copy and have enough space. --verify reads every file to validate SHA256; it can take substantial time. Never point restoration at an existing populated directory.

Recorded evaluation

The recorded four-window teacher comparison reports KLD 0.0485539331 nats and top-1 agreement 92.6844%. These are limited calibration-window measurements, not a broad intelligence benchmark, not a percentage of information retained, and not comparable to unrelated GGUF evaluations without matched protocols. See the source repository for runtime/evaluation details and repeatability limits.

The qualified K2.75 profile measured 13.062 decode tok/s over six llama-benchy runs with pp2048/requested tg256, concurrency 1. A 30,039-token retrieval prompt passed. The profile uses resident UVA, TP3/DCP3, MTP3/draft TP1, eager, prefill128, NCCL4/1MiB, FP8 KV 1GiB/rank, 32,768 total context, and a 12GiB host reserve.

Provenance and integrity

• BF16 source revision: 304b8051cfb2b260b61ce0cbe330e02a98e73639.
• Sealed source directory: GLM-5.3-EXL3-TR3-2.75bpw-TP3-K2K3-rotating-uneven-v1-assembled-20260907T210500Z.
• Original manifest SHA256: d0384bfd53e403b859e8a221b0429725b22600d48d49f2ada2cf8a0f777332c6.
• Original checksum-ledger SHA256: b3311519773b35a8ede1fe10975df5e97096e1870fc994c51626a1da03ea2457.
• Full NAS archive audit on September 8 checked both checkpoints, 39,266 ledger entries in total.
• Publication verifies remote file sizes and content identities against the transport manifest before adding UPLOAD_COMPLETE.json.

Original model licensing is preserved in LICENSE. This repository does not replace that license with MIT or Apache. Quantized weight bytes are unchanged from the sealed NAS checkpoint.