Modelpublic

davidsyoung/GLM-5.3-EXL3-TR3-3.25bpw

sourceHugging Faceotherupdated 23d agoView on Hugging Face

7likes1.6kdownloads

Model Card

GLM-5.3 — EXL3/TR3 3.25 bpw (mixed K3/K4 trellis, data-free)

[!NOTE] Weights uploaded and verified; KLD measured (2026-08-29, full-vocabulary teacher-forced KLD vs the sealed BF16 teacher, held-out confirmation windows of brandonmusic/GLM-5.3-BF16-full-logits — 4 windows x 2,047 positions x 154,880 vocab; reproduction kit + runner in this repo).

Trellis (EXL3) quantization of zai-org/GLM-5.3 (755B glmmoedsa MoE: 78 layers + MTP, 256 routed experts/layer), following the GLM-5.2 TR3 lineage layout. Fits TP4 on 4x 96 GB (RTX PRO 6000 Blackwell class) with FP8 KV cache.

From the upstream card: "GLM-5.3 uses the same base model as GLM-5.2 — every gain comes from post-training. Compared with GLM-5.2, it is much better at complex coding and long-horizon tasks." See the base model card for benchmarks, serving guides, and the technical report (arXiv:2602.15763).

What is quantized

component	treatment
routed experts (layers 3–78, incl. MTP-78)	EXL3 trellis, per layer 192 experts K3 + 64 experts K4 (avg 3.25 bpw), mcg codebook
dense MLP (layers 0–2), all attention, norms, embeddings, lmhead, mlp.gate, ehproj	BF16, carried byte-exact
shared experts	BF16 in-checkpoint (online K6 at serve)

—Data-free encode: identity Hessian (H=I), rotations + trellis search only — no calibration capture anywhere. Deterministic: seeds derived from (layer, expert, projection, rank), seed_base 20260711.
—K4 selection: per layer, the experts with highest relative round-trip MSE under the K3 encode are re-encoded at K4 (worst-64 per layer).
—Per-expert tier map in tier_bitmap.json; encode provenance in config.json.hybrid_tr3_tail; file hashes in MANIFEST.sha256.

Serving

TR3 mixed-bit layout (carrier BF16 shards + per-layer trellis payloads) — serve with an exllamav3-b12x/sparkinfer-lineage stack, TP4 (+ DCP4/MTP3), FP8 KV. The mixed-K projection-tiers patch is REQUIRED; a stock loader that assumes a uniform K per layer will produce fluent garbage. Not loadable by vanilla exllamav3 model loading.

KLD vs BF16 teacher

Full-vocabulary, teacher-forced KL(teacher || student) against sealed BF16 GLM-5.3 logits — 4 held-out windows × 2,047 positions × 154,880 vocabulary, fp32 log-softmax both sides. Measured independently on two different 4× RTX PRO 6000 (96GB) machines:

weight quant	KV mode	this work	CN3 ([@dareposte](https://huggingface.co/dareposte))	Δ
3.42 bpw	fp8	0.024105	0.023966	−0.6%
3.25 bpw	fp8	0.026103	0.026776	+2.6%
3.25 bpw	nvfp4	0.035741	0.036661	+2.6%
3.42 bpw	nvfp4	0.037757	0.037060	−1.8%
3.42 bpw	nvfp4+rope8	0.039518	0.037695	−4.6%
3.25 bpw	nvfp4+rope8	—	0.039396	CN3 only

Readings: the 3.25↔3.42 weight step changes KLD by only ~0.002; the fp8→nvfp4 KV step costs ~7× more (~0.014) — cache format matters more than the extra 0.17 bpw. Under nvfp4 KV the weight-quant difference washes out entirely. The window SD (~0.02–0.03) is corpus heterogeneity — one citation-dense legal window is uniformly hardest; dialogue, explanatory prose, and reasoning-trace registers measure near-transparent.

<details> <summary>Per-window means (dialogue / legal / prose / reasoning-trace)</summary>

config	source	w0000	w0001	w0002	w0003
3.42 fp8	this work	0.0141	0.0537	0.0137	0.0148
3.42 fp8	CN3	0.0148	0.0542	0.0135	0.0134
3.25 fp8	this work	0.0188	0.0579	0.0138	0.0139
3.25 fp8	CN3	0.0183	0.0592	0.0142	0.0154
3.42 nvfp4	this work	0.0199	0.0816	0.0221	0.0274
3.42 nvfp4	CN3	0.0198	0.0775	0.0210	0.0300
3.25 nvfp4	this work	0.0256	0.0799	0.0184	0.0190
3.25 nvfp4	CN3	0.0255	0.0804	0.0189	0.0218
3.42 nvfp4+rope8	this work	0.0212	0.0812	0.0257	0.0300
3.42 nvfp4+rope8	CN3	0.0218	0.0761	0.0236	0.0292
3.25 nvfp4+rope8	CN3	0.0268	0.0883	0.0204	0.0221

</details>

Method (reproducible)

—Teacher: brandonmusic/GLM-5.3-BF16-full-logits, reference-full-panel confirmation lane (held out from every calibration fit), revision 427368f1.
—Student: this checkpoint, loaded by the digest-pinned r17 serving image (sha256:c5e96c5b…) — the real trellis kernels and online-K6 path, offline vllm.LLM, TP4, one teacher-forced prefill per window.
—Full runbook + runner: `kld/` in this repo (KLD-REPRODUCTION.md, prefill_kld_53.py, fetch-teacher.sh).
—Independent reproduction bundle (receipts, unedited logs, checksums, pinned revisions): `kld/cn3/`.

Capacity caveat (via the CN3 report): KV-pool token figures printed by KLD-profile boots (TP4/DCP1, 4,096-token envelope) are logical pool values for that profile only — they are not maximum context length or production serving capacity.

Credits

—[brandonmusic](https://huggingface.co/brandonmusic) — thank you for the GLM-5.3-BF16-full-logits teacher captures that make this measurement possible without a 1.5TB BF16 forward, and for the TR3 quantization references this release follows: the GLM-5.2-EXL3-TR3v4-3.5bpw-MTP78 method card, runbook, and the r10 reproducibility bundle whose encoder lineage (encode_tr3_v31.py) this checkpoint was produced with.
—[dareposte](https://huggingface.co/dareposte) — thank you for the independent CN3 reproduction: all six weight/KV configurations on separate hardware, within ±5% of our means, published here with receipts under `kld/cn3/`.
—willfalco/GLM-5.2-EXL3-TR3-3.25bpw — the 3.25 tier recipe and mixed-K checkpoint-format lineage.
—local-inference-lab — the qualified r17 serving stack these artifacts boot on.

Status

Measured and independently reproduced 2026-08-29: mean KLD 0.026103 (fp8 KV; CN3 reproduction 0.026776) on held-out confirmation windows. Serving-qualified on the r17 stack; measured 590k-token KV pool at GMU 0.965 — the context-breadth artifact of this release pair.

Quantized with encode_tr3_53.py (exllamav3 v0.0.43 vendored math, MIT) on 4x RTX PRO 6000 Blackwell. Encode + tooling notes ship in the repo.