base_model:

• Mia-AiLab/GLM-5.3-Flash-EXL3-TR3-4bpw
• dealignai/GLM-5.3-Flash-UNCENSORED-NVFP4
• wrldsuksgo2mars/GLM-5.3-Flash-EXL3-K3-v1 base_model_relation: merge pipeline_tag: image-text-to-text library_name: transformers license: other license_name: shapleymcg-license-1.0 license_link: https://github.com/brandonmmusic-max/shapleymcg/blob/main/LICENSE tags:
• glm5
• glm5-next
• mixture-of-experts
• exl3
• tr3
• mixed-precision
• quantized
• uncensored
• 1m-context
• dgx-spark
• shapleymcg
• mirror

GLM-5.3 Flash Uncensored EXL3/TR3 3.51 bpw (mirror)

Mirror notice. This repository is an unmodified mirror of s-zaizen/GLM-5.3-Flash-EXL3-TR3-3.51bpw-Uncensored at revision 6dff879f2cb6784c3c10401e5ecd6b6c483ee121 (the source repository no longer resolves publicly as of 2026-10-31). All 120 weight shards, configs, tokenizer, provenance receipts, LICENSE, ATTRIBUTION and THIRD_PARTY_LICENSES files are byte-identical to that revision. Only this README has been edited: the notice above and the deployment notes at the end were added, and the original model card text is kept unchanged below. All credit belongs to the original authors listed under License and attribution. The mirror maintainer claims no authorship of the model, the quantization, or the build.

This work includes or was produced using ShapleyMcg, created by Brandon M. Music (https://github.com/brandonmmusic-max/shapleymcg). ShapleyMcg is licensed under the ShapleyMcg License v1.0, an attribution-required license that grants no rights to the person known as "0xSero." Use of ShapleyMcg without this attribution is unlicensed.

This checkpoint is a reproducible, layer-mixed K3/K4 derivative intended for two NVIDIA DGX Spark nodes. Its main goal is to create enough unified-memory headroom for the native 1,048,576-token context while retaining the sensitive parts of the existing K4 checkpoint.

Publication identity:

• local artifact revision: 024db9-e90ef-180a1f-layer-mix-v1;
• artifact-content SHA-256: 95c9d9d50122c886fc2e9c18152b5d81a457a85d1a20e2dc5f57e57d676c707e;
• 120 model shards;
• effective routed-expert rate: 3.511627906976744 bpw.

What changed

• All attention, linear-attention/KDA, DSA indexers, mHC, routers, shared experts, dense layers, embeddings, LM head, norms, vision, and tokenizer assets remain byte-identical to the local uncensored K4 carrier.
• BF16 attention output projections in layers 15–44 use the pinned uncensored donor values. Layer 45 was already donor-identical in the carrier.
• Routed experts in 21 complete layers use the pinned public K3 payloads.
• Routed experts in the other 22 layers, including layers 3–5 and 42–45, stay K4. The effective routed-expert rate is 3.511627906976744 bpw.
• No training, SFT, DPO, LoRA merge, dequantization, or re-quantization is performed by this build. Packed tensor payloads are copied byte-for-byte.

Complete-layer mixing is required by the current fused EXL3 runtime, which uses one trellis width per MoE layer. The chosen K3 layers minimize the sum of the published Hessian-weighted K3 projection error multiplied by routed sample count, while boundary layers and the checkpoint MTP layer remain K4.

Reproducibility

The artifact includes:

• MIXED_EXL3_ALLOCATION.json — the fixed layer-rate plan and per-layer input measurements;
• K3_SOURCE_PROVENANCE.json — the K3 publisher's source and quantization receipt;
• K3_QUANT_LOG.csv, K3_VALIDATION_REPORT.json, and K3_ERROR_LEDGER_MANIFEST.json — the pinned selection evidence copied from the public K3 source;
• MIXED_EXL3_BUILD.json — source revisions, every output shard hash, and the aggregate artifact identity;
• DERIVATIVE_PROVENANCE.md — scope and limits of the final mixed artifact;
• K4_UNCENSORED_DERIVATIVE_PROVENANCE.md and UNCENSORED_BUILD.json — the intermediate uncensored K4 carrier's provenance and byte-transplant receipt;
• the original attribution and ShapleyMCG license documents from the K4 carrier.

Pinned inputs:

• K4 carrier: Mia-AiLab/GLM-5.3-Flash-EXL3-TR3-4bpw at 024db9f7e9871e8efdf21538ba55af7442be3cd5;
• uncensored donor: dealignai/GLM-5.3-Flash-UNCENSORED-NVFP4 at e90ef415a2bf1274f848965a61144eca1e99eabf;
• K3 payload source: wrldsuksgo2mars/GLM-5.3-Flash-EXL3-K3-v1 at 180a1f118595725491b29e6e1a24785622371a64;
• common BF16 source index SHA-256: e6007bd58fb7e07f9fe69544257ee2713f252ef5855bbf685b48c991d524ef0f.

Runtime requirement

Stock vLLM does not currently implement this checkpoint's GLM/EXL3 reader. The qualified runtime is the layer-mixed profile in s-zaizen/dgx-memo@ad7fafc, based on MiaAI-Lab's two-node runtime at commit 79f10b91f84779b2b1ff2c9327b1a5847cd97f70 and base image digest sha256:9bb1557a4234fce63d59599e44d10747eabd742beb337eebf9e7070be8a0fd58. Its small reader overlay adds complete-layer layer_bits selection; each layer still uses the existing uniform-rate fused EXL3 kernel. The model's quantization_config.json records the per-layer rates.

The exact qualified profile uses two DGX Spark nodes and these material settings:

TP=2
MAX_MODEL_LEN=1048576
MAX_NUM_SEQS=1
MAX_NUM_BATCHED_TOKENS=2048
KV_CACHE_DTYPE=fp8
KV_CACHE_MEMORY_BYTES=12884901888
SPEC_METHOD=dflash
DFLASH_MODEL=incoai/GLM-5.3-Flash-DFlash2
DFLASH_TOKENS=7

The DFlash2 checkpoint is fetched separately and is not included here. It is published under CC-BY-NC-ND-4.0; review that license before enabling speculative decoding. DFlash proposes tokens only, and the target model verifies every accepted token.

That deployment script is designed for a two-node DGX Spark installation; it is not a stock-vLLM launch command. Keep substantial host-memory headroom and qualify progressively before attempting the full context.

Evaluation status

The exact artifact identity above was evaluated on 2026-08-30 with TP=2, DFlash2 K=7, FP8 KV, a 12 GiB KV allocation per rank, one active sequence, and a 1,048,576-token configured maximum.

The cold isolated 1M run had a 0% prefix-cache hit rate. The long suite checked exact-marker retrieval and ran a short health probe after every long request. No target runtime errors were found in the final head or worker logs.

These are system-specific qualification results, not general capability benchmarks. Throughput and memory headroom depend strongly on host services, interconnect, runtime revision, KV allocation, and prefix-cache state. The qualified profile has concurrency one at the full context, and the cold 1M prefill takes about 19 minutes. No capability or quality score is inherited from any parent checkpoint.

This checkpoint is uncensored/abliterated and may produce harmful, inaccurate, or policy-violating content. It is not suitable for unsupervised deployment. Operators remain responsible for access control, downstream safeguards, and applicable law.

License and attribution

The K4 carrier is subject to the ShapleyMCG License 1.0. The K3 source declares MIT licensing, and the base GLM model is published by Z.AI. Keep all bundled license, attribution, provenance, and third-party notice files when copying or redistributing this artifact.

In particular, retain the complete ShapleyMCG License 1.0, the required attribution notice, the final derivative provenance, the build receipt, and all files under THIRD_PARTY_LICENSES/. The named exclusion in Sections 4–6 of the ShapleyMCG License remains in force for this derivative; this model is source-available and must not be described as OSI open source.

Mirror maintainer's deployment notes (2× DGX Spark)

Independent measurements by the mirror maintainer, not part of the original qualification. Setup: two DGX Spark (GB10, 121.63 GiB unified memory each), TP=2 over ConnectX-7 RoCE, MiaAI-Lab kit overlay with the layer-mixed EXL3 reader, vLLM Model Runner V2, FP8 KV, 12 GiB KV pin per rank, DFlash2 K=7, one sequence, 2,048 batched tokens. Measured in September 2026.

Prefill is roughly half of what the kit reports with its defaults (EXL3_FAT_KERNEL=1, MAX_NUM_BATCHED_TOKENS=7168). Those were left off because the fat kernel was untested at the 3-bit layers. Deployment scripts, benchmarks and reports: github.com/baonq-me/glm-5.3-flash-exl3-tr3-3.51bpw-uncensored.