GLM-5.3-Flash selective EXL3 3.0 bpw

[!WARNING] This checkpoint uses a custom selective-EXL3 TP=4 layout. It has not passed full-server loading, endpoint, or generated-response acceptance. Vision and MTP execution are also unvalidated. The held-out quality result was measured successfully but fell outside the conservative Q4 control gates described below.

This is a selective 3.0 bpw EXL3 conversion of Z.AI's GLM-5.3-Flash-BF16, pinned to source revision a6c167b62691b2bac901344b65cb651a70f53e43. Compatibility with the later official source revision f12e0fe1f6b2ea274c11a569582edfd99d993c5e was separately checked before release assembly.

Only routed-expert gate/up/down projections in language layers 3–44 are EXL3 K3. Attention, linear attention, indexers, mHC, routers, shared experts, dense layers 0–2, embeddings, LM head, norms, vision, and MTP remain at source BF16.

Status

Layout and size

The artifact contains 583,090 indexed tensors and 149.56 GB of files. Routed-expert EXL3 tensors account for 115.49 GB; 2,482 retained source-precision tensors account for 33.84 GB. The format identifier is glm53-selective-exl3-tp4-v1, and a compatible custom loader is required. Stock Transformers and generic EXL3 compatibility are not claimed.

Calibration and coverage

Conversion used 600 sealed rows × 2,048 tokens = 1,228,800 tokens from the pinned ExLlamaV3 calibration bundle plus bounded synthetic coverage rows. Natural top-8 routing covered every expert in every routed layer. Calibration was used only for conversion; this release does not train or fine-tune the base model.

The calibration-row SHA-256 is 1cae9bbcd2beb3879a0c459edfca1fd197043ab204b82189c9361de386d0cae1. The conversion used ExLlamaV3 commit 5f3c537ca9d89893d771256f5c43c93656553fbb.

Validation results

The scoped runtime primitive test loaded real EXL3 gate/up/down tensors for all four TP ranks in two waves on two physical GPUs. Every rank captured and replayed a CUDA graph, with zero graph-replay-versus-eager relative L2 error in that test. This does not prove collectives or complete model serving.

Held-out evaluation used 65,504 next-token positions from 32 contiguous 2,048-token blocks of the Salesforce Wikitext-2 raw test split, disjoint from calibration rows.

The raw reports are under evidence/. The release preserves this result without relabeling it as a quality pass.

Sibling releases

• EXL3 suite
• 2.5 bpw — pending
• 2.0 bpw — pending
• 4.0 bpw control

Attribution and license

• Z.AI: base model and MIT license.
• Dione: selective EXL3 conversion and validation workflow.
• ExLlamaV3 / TurboDerp: EXL3 format and conversion implementation.
• Brandon M. Music: earlier MIT-licensed GLM EXL3/TR3 rank-sliced lineage and the public release-engineering pattern of reproducibility closure, cold KLD receipts, explicit runtime profiles, and content-addressed provenance. See brandonmusic/GLM-5.2-EXL3-TR3-3.0bpw and the separately licensed GLM-5.3 release. No ShapleyMCG code, model weights, calibration corpus, or generated artifact is included or used by this release.
• Salesforce Research: Wikitext evaluation dataset.

The source LICENSE is included. Use this checkpoint only with a loader that understands its selective GLM-5.3 EXL3 TP=4 contract, and treat full serving, vision, and MTP support as pending until separately demonstrated.

Reproducibility, results, runtime, and provenance

• reproducibility/ records the public-safe source-closure status and verifier.
• results/ indexes raw held-out quality receipts without hiding the failed conservative gate.
• runtime/ separates scoped kernel/CUDA-graph evidence from the still-pending full-server and MTP gates.
• PROVENANCE.md defines the content-addressed identity chain.
• RELEASE_STATUS.json is the machine-readable gate ledger for this exact Hub snapshot.

Release identity

Status: Public checkpoint. Audited weight payload: 130 safetensors files, 149,402,871,912 bytes (weight files only; excludes metadata).

Upstream source: zai-org/GLM-5.3-Flash-BF16, BF16 revision a6c167b62691b2bac901344b65cb651a70f53e43. Artifact/evidence snapshot inspected: 8b099bf276507a17faea920deff3f62d5597fb52. A card update does not constitute a new weight conversion.

Intended use and limitations

Use populated checkpoints for local inference or quantization research with the declared compatible runtime. Results from one bitrate or runtime do not transfer automatically to another. Quantization may change behavior and factual accuracy; controlled smoke tests do not establish broad benchmark quality. A projection or REAP observation record alone does not prove successful refusal removal or a pruned model release.

Related releases

Private links require authorized access. Two suite indexes and three placeholders are included in this inventory; they are not additional trained models.

Evidence files

• CALIBRATION_COVERAGE.json
• EXL3_MANIFEST.json
• evidence/quality-report.json
• evidence/release-calibration-manifest.json
• evidence/route-coverage-manifest.json
• evidence/runtime-status.json

REAP observation provenance

The original 3bpw and Q4 were each observed on two corpora: private calibration material and balanced 12-language Wikipedia. Each sealed lane records 128 sequences × 1,024 tokens (131,072 tokens), across 42 routed layers and 288 experts per layer. These four observation lanes are separate from quantization calibration and do not mean experts have been pruned from the weights above. The private observation dataset holds the manifests and aggregate sidecars.