GLM-5.3 (full) on two DGX Sparks: mixed-K EXL3

Built on Z.ai's GLM-5.3 and turboderp's EXL3 format. This is an independent community release; none of the people credited here have reviewed or endorsed it.

The full GLM-5.3 (753B parameters, mixture of experts) quantized to run on two NVIDIA DGX Spark machines, with 262k context and every expert kept.


Base model	zai-org/GLM-5.3, FP8 release
Parameters	753B total, about 40B active per token (top 8 of 256 experts)
Routed experts	EXL3 trellis, mixed widths, ~2.0 bits per weight
Attention, shared expert, dense layers	FP8 weights, as released
Embeddings and output head	BF16, as released
Experts	all 256 per layer, none pruned or merged
Built-in MTP draft layer	not included
Size on disk	203.3 GB
Context	262,144 tokens
Target hardware	2x DGX Spark (GB10)

How close is it to the original?

Short answer: very close where the original is sure, and close overall, but not lossless. Expect some replies to differ.

Score	What it tells you	Result
Generated-token KLD, held-out text	How far the predictions differ on tokens a model writes, on mixed text.	0.033
Generated-token KLD, long replies	The same, on the original's own long replies. This is the harder test.	0.122
Top-1 agreement, held-out text	How often both models pick the same next token.	93.3%
Top-1 agreement, long replies	The same, on long replies.	89.3%
Confident-token agreement	How often they match where the original is sure of the next token.	97-98%
Tool-call KLD	How far the predictions differ on structured tool calls.	0.021
All-token KLD	A text exactness check: every token, including text a model reads but never writes.	0.282

Behaviour. The biggest win is coherence. In my testing so far, replies stayed coherent with no garbled output, 18 of 18 agent tool calls were valid JSON, and 7 of 7 strict tool checks passed. This is a small test, not a benchmark.

Decision score, the original's own replies	What it tells you	Result
Hesitation ratio	How much of its prediction goes to hesitation words, against the original. 1.0 means the same.	0.98
Tool-call decisions	KLD at the points where it chooses to call a tool.	0.002
Ending its thinking	KLD at the point where reasoning closes.	0.079
Starting the answer	KLD at the first answer token.	0.111
Ending a reply	KLD at the end of a turn. This is the weakest score here, on few positions, and the tokens that end a turn were not split apart.	0.422

These come from the 2k, 4k and 8k long-reply buckets.

What these numbers mean

I gave this model and the original the same text. At each position, both predict the next token, and I compared the two predictions. Neither model writes freely in this test, so the differences come from the models and not from a different reply.

KLD measures how far apart two predictions are. 0 means identical, and lower is closer. Think of it as how surprised the original would be by this model's odds.
Top-1 agreement counts how often both models rank the same token first. At 93.3%, about 1 token in 15 differs. At 89.3%, about 1 in 9 differs. A different top token can be a synonym or another valid choice, so 1 token in 15 differing does not mean 1 in 15 is wrong.
Confident-token agreement looks only at positions where the original gives its top token more than 50%. A flip here can change an answer, so this score matters most. At 97-98%, about 1 in 40 differs.

Why two test sets? Held-out text is a broad mix of code, maths, science, tool use and prose. Long replies are 117 replies written by the original, with long reasoning, for 164,074 generated tokens in total. Its scores are lower, and I show both because a broad test alone would flatter this quant. The held-out windows were also used during tuning, so treat 0.033 and 93.3% as somewhat optimistic.

Why two kinds of KLD? Generated-token KLD counts only what a model writes: reasoning, answers and tool calls, so it is the closest score to output quality. All-token KLD is a token and text exactness check. It also counts prompts, tool schemas and tool results, which a model reads and never writes, so it reads much higher.

"The original" means Z.ai's FP8 weights run with BF16 activations. Confident-token agreement comes from the four task sets below, and all-token KLD comes from general text. Across all five test sets, all-token KLD is 0.283 and top-1 is 85.2%.

Every score was measured on the exact weights in this repository, offline, with a reference evaluator. I have not yet repeated them through a live server, so served numbers may differ slightly. The paired check further down did run through a live server.

By task (reasoning on)

These rows use separate test sets from the first two rows above, so they do not average to 0.033.

Task	What it covers	Generated-token KLD	All-token KLD	Confident-token agreement
Code	writing and editing code	0.059	0.201	97.9%
Maths	step-by-step problem solving	0.031	0.039	97.8%
Science	expert science reasoning	0.048	0.081	97.4%
Tool use	turns with tool calls	0.033	0.811	97.6%

Tool use has a high all-token KLD because that count includes tool schemas and tool results, which the model never writes.

For scale: the original against itself

Two runs of the same weights are not identical when their numerics differ. I ran the original weights with FP8 activations and compared them with the same weights on BF16 activations.

	All-token KLD, general text	Top-1 agreement
The original, FP8 vs BF16 activations	0.019	96.5%
This model vs the original	0.282	see above

I have not measured this floor on generated tokens, so I do not state a multiple for the 0.033.

Benchmarks, side by side with the original

An early paired check through a live server, at 262k context, on the same questions as my API runs of the original. Both models used T=1, top_p 0.95 and maximum effort.

Check	This model	The original
MMLU-Pro, 12 questions	9 right	9 right
GPQA-D, 12 questions	9 right	9.0 right, average across runs
Thinking loops, unclosed thinking, cut-off replies	0 of 24	Cut-offs 0 of 24, no text to compare the rest
Total tokens	330,456	about 341,200

This is a small sample, chosen to include hard questions, so its 75% is not a benchmark score. The 95% range on the accuracy gap is about -8 to +11 points, so it shows no sign of a drop and does not prove there is none. Typical reasoning ran about 11% longer, while total tokens were about 3% lower. A larger run will follow.

How to run

These packs use per-matrix EXL3 rotations, so they do not load in standard EXL3, EXLR8 or gr8 runtimes. Run them with the recipe: Kristianaaron/GLM-5.3-Mixed-K-EXL3-2x-DGX-Sparks.

Licence

These weights are a quantization of Z.ai's GLM-5.3 and a derivative work under the GLM-5.3 License, Copyright (c) 2026 Z.AI. The full licence text and its notice are in LICENSE. Its conditions apply to this release, including the security review condition for large Model as a Service operators. The weights are provided as is, with no warranty.

Credits

I built this release on the work of the people below. Thank you.

Special thanks

Matt Mastracci (mmastrac, Kindling AI), for EXLR8. EXLR8 showed how far mixed-width quantization of GLM-5.3 could go, and its multi-width store and FP8-aware rounding ideas directly shaped this pack. Please read the EXLR8 model card.

The model and the format

Who	Role in this release	Licence	Source
Z.ai	GLM-5.3, the base model. Technical report: GLM-5: from Vibe Coding to Agentic Engineering	GLM-5.3 License	zai-org/GLM-5.3
turboderp	ExLlamaV3: the EXL3 format, trellis codec and encoder. I encoded with exllamav3 1.5.4	MIT, Copyright (c) 2025 Turboderp	turboderp-org/exllamav3
Matt Mastracci and Kindling AI	EXLR8: the multi-width store and FP8-aware rounding ideas behind this mixed-width pack	Format additions CC BY 4.0, on top of the GLM-5.3 License	kindlingai/exlr8, model card
Tseng, Sun, Hou, De Sa	QTIP, the trellis-coded quantization paper behind EXL3	paper	arXiv:2406.11235
Tseng, Chee, Sun, Kuleshov, De Sa; Chee, Cai, Kuleshov, De Sa	QuIP# and QuIP, incoherence processing	papers	arXiv:2402.04396, arXiv:2307.13304