"GLM-5.3-class": Scantist fires a quarter of the experts, shows 92.7%, and doesn't put the original through the same test

In short

• Since 02/11/2026, Scantist has published a version of GLM-5.3 that keeps 192 of the 256 experts in each layer, picked by an automated search, with no retraining. It weighs 179.86 GB, versus 238.58 GB for the Unsloth compressed version it starts from.
• According to Scantist, it fits on a single NVIDIA B200, or on a Mac Studio with at least 192 GB of unified memory. Its guide and its measurements only cover the B200: nothing is measured on a Mac.
• Its scores, 92.7% on HumanEval (coding exercises) and 84.8% on GPQA-Diamond (a science multiple-choice test), have nothing to compare against: in its own table, the 256-expert version is marked "not measured".

Sunday evening: who put GLM-5.3 on a diet?

Logbook. I open the card. It promises "a GLM-5.3-class model" on a single graphics card. No, seriously. GLM-5.3 is Z.AI's big model, which the lab calls "the most capable open-weights model for coding". It has 753 billion parameters, the settings learned during training, and its weights take about 755 GB in the version Z.AI distributes, by our calculation.

First reflex: who signed it? Not Z.AI. The repository belongs to Scantist, and the card credits everyone: weights from Z.AI, compression from Unsloth, expert selection from another project, GLM-5.3-SLIM-E192. Scantist puts it together. On Hugging Face's counter, on 08/11 around 19:40 UTC: 4,938 downloads and 949 likes.

The starting point is already a compressed version of GLM-5.3, published by Unsloth in GGUF, a file format for compressed models. According to Scantist, it weighs 238.58 GB and fits neither in the memory of one B200 nor in a 192 GB Mac. So Scantist cuts.

What is an expert, and why fire 64 per layer?

GLM-5.3 is a mixture of experts. In 75 of its computing layers, 256 small sub-networks, the experts, wait their turn. A switchman, the router, picks 8 of them for each token, each word fragment the model reads or writes.

Picture a gacha game. You have 256 characters in your box, you only field 8 per fight, and the game picks them based on the enemy. Scantist emptied 64 characters out of every box. If you never used them, you won't notice. If that was your ice-boss squad, you will.

In practice, each of the 75 expert layers keeps the 192 chosen by an automated architecture search; the other 64 are gone. Our calculation: 4,800 experts removed out of 19,200. How did the search choose? The card points to the GLM-5.3-SLIM project, under the autotrust account. Its page answered us with a "401", access denied: we did not read it. Method not detailed, then.

According to the card, the rest is clean work. The kept experts are copied "byte-for-byte, in their original order". Everything else in the model is identical to Unsloth's file, except a side layer used to generate faster, called MTP, which is removed too. "No retraining, no recalibration". Scantist says it compared its 1,782 tensors, the blocks of numbers in the file, with those of the source. Result: 563 billion parameters instead of 753, and a file 24.6% lighter than Unsloth's.

2.56 bits per weight: the Game Boy edition of GLM-5.3

Second diet, an older one: quantization. Each weight in the model is a number. Z.AI stores almost all of its weights on 8 bits. With 8 bits you can write 256 different values. With 2 bits, four. With 3, eight.

Scantist's file comes down to 2.56 bits per weight on average. An average, because not everything gets the same treatment: the experts are stored in the tightest formats, mostly from the IQ2 and IQ3 families, while attention, the part that links words together, keeps wider formats. It's a bit like going from a colour photo to a four-shade image, old handheld-console style: you still recognise the scene, you lose the nuances.

Sources: Scantist and Z.AI cards, Hugging Face data; size of Z.AI's version and B200 capacity: our calculation.

From 755 to 180 GB, the model is about 4.2 times lighter (our calculation). If you download it at 1 gigabit per second without interruption, count about 24 minutes. Scantist sums it up itself: "Two losses stack: 25% of the routed experts are removed, and the weights are at ~2.6 bits per weight".

The gap: 92.7%, compared with what?

Right. The scores. Scantist publishes three, measured with its own scripts on a B200:

• HumanEval: 164 small programming exercises. The model writes a program from a description, a script runs it against the official tests. 92.7%, meaning 152 exercises solved on the first try, with thinking turned off.
• GPQA-Diamond: 198 multiple-choice questions in biology, physics and chemistry, written by specialists to be "extremely difficult". 84.8%, meaning 168 right answers, with high reasoning effort and up to 65,536 tokens per answer.
• MMLU: a big knowledge test covering 57 subjects, from elementary maths to law. 71.4%, on a 570-question subset, ten per subject.

Wait. In the next column, the one for the 256-expert version, it says "not measured". And Scantist writes it in bold further down: "We did not run the original 256-expert UD-IQ2_M or the full-precision model on the same harness, so the cost of the expert reduction itself is not measured here". And the official GLM-5.3 card is no reference point: Scantist warns that its numbers "are not comparable to the numbers in the official GLM-5.3 card", which has neither HumanEval nor GPQA-Diamond anyway.

It's a speedrun time with no reference time. If the full version did clearly better, the cut cost a lot. If it did the same, it cost almost nothing. We don't know which, and Scantist didn't measure it either: it says so itself.

The only clue, Scantist gives it in one line: the before/after tests of the SLIM project, on the same expert set, "show a clear loss on Chinese knowledge QA and a smaller loss on science QA; expect the same here". For the numbers, Scantist points to the autotrust page, the one that shut the door on us. Maybe my ice-boss squad spoke Chinese.

When NVIDIA moved DeepSeek to another 4-bit format, its card put the original and the converted version side by side, on six tests. That's the column missing from Scantist's card.

Promised on a Mac, measured only on a B200

Picture it. You have a Mac Studio with 192 GB of memory or more on your desk, or you dream of one, and you read that "a GLM-5.3-class model" will fit in it. Scantist's guide, though, ran everything on a single B200, a data-centre graphics card whose reported memory comes to about 192 GB (our calculation). The Mac is named in the promise, with the mlx-serve tool, but I find no measurement and no instructions for it in the card.

On the B200, the weights alone take about 167 GiB, close to 180 GB (our calculation). That leaves room for about 100,000 tokens of working memory, shared by up to 10 conversations. Speed measured by Scantist: 34.5 tokens per second for a single conversation, within about 10% from one run to the next. A 500-token answer would take about 15 seconds (our calculation). Their full GPQA-Diamond run, on the other hand, took about 13.5 hours: to keep the working memory from overflowing, the script had to abort and restart requests 275 times.

And on a PC without a graphics card? No luck either. With the tested version of llama.cpp, the engine cannot run this architecture on the processor alone, and neither can the original 256-expert version.

No B200 on our side: everything I'm telling you comes from the card. Kira has already taken apart the other end of this kind of hack, on another mixture of experts: Kimi K3 in 8 GB of RAM, at the cost of 1.7 TB of disk.

The GLM-5.3 licence: almost everything allowed, except for very large services

Scantist publishes under the GLM-5.3 licence, "the same terms as the base model". I read Z.AI's text in full. It's short.

What it allows, free of charge: using, copying, modifying, publishing, distributing, selling and fine-tuning the model, and creating derivative works from it.

What it requires:

• keep the copyright notice and the licence text in all copies;
• comply with applicable laws and regulations;
• a clause for the giants, quoted word for word: "If the Licensee or any of its affiliates operates a Model as a Service business, and the aggregate revenue of the Licensee and its affiliates exceeds 10 billion US dollars (or the equivalent in other currencies) in total over any consecutive 12 months, the Licensee must pass Z.AI's security review before using the Software or its derivative works for any commercial purpose."

The licence defines "Model as a Service": giving a third party access to the model, through an API (a doorway for software) for example, with meaningful control over the inputs, parameters or training data. An app that embeds the model in a specific feature is not one. Everything is provided "as is", without warranty. Nowhere does the text call itself an open-source licence; Hugging Face files it under "other", and Z.AI talks about open weights.

What we don't know yet

• What the cut loses. Not measured in Scantist's repository; it points to the SLIM project's before/after tests, on a page that denied us access.
• How the 64 experts were chosen. The card says "neural architecture search", without detail.
• An outside opinion. No independent test of this version in our sources.

My verdict: the cut is clean work, checked tensor by tensor according to Scantist, and fitting 563 billion parameters onto a single graphics card is top-flight tinkering. But "GLM-5.3-class" without GLM-5.3 on the same bench is a slogan, not a measurement. And the loss figures it points to sit behind a page that answered us "401".

Key terms

• Mixture of experts: a model split into many sub-networks, the experts, only a few of which work on each token.
• Router: the part of the model that picks which experts work on a given token.
• Parameter: one of the numbers tuned during training. GLM-5.3 has 753 billion.
• Token: a word fragment, the unit a model reads and writes.
• Quantization: storing each weight on fewer bits to make the model lighter, at the cost of some precision.
• Bits per weight: the space each number in the model takes. 2 bits allow 4 values, 8 bits allow 256.
• GGUF: a file format for compressed models, read in particular by llama.cpp, a program that runs models on your own machine.
• Neural architecture search (NAS): an automated search for the model's shape; for GLM-5.3-SLIM, it set which experts to keep.
• MTP: a side layer of GLM-5.3 used for speculative decoding, a method to generate faster; removed in Scantist's version.
• B200: an NVIDIA graphics card for data centres, with about 192 GB of memory.
• Unified memory: the name Apple gives to the memory of its Mac Studio.
• HumanEval: 164 programming exercises, scored by running the model's code against tests.
• GPQA-Diamond: 198 multiple-choice questions in biology, physics and chemistry, written to be extremely difficult.
• MMLU: a knowledge test covering 57 subjects, from elementary maths to law.

Source sheet

• Scantist's repository: Scantist/GLM-5.3-SLIM-GGUF, created on 02/11/2026, last modified on 03/11/2026, GLM-5.3 licence.
• Original model: zai-org/GLM-5.3, Z.AI, card created on 25/08/2026, GLM-5.3 licence.
• Compressed starting version: unsloth/GLM-5.3-GGUF, UD-IQ2_M folder.
• Expert-selection project: autotrust/GLM-5.3-SLIM, 401 response on 08/11/2026, not read.
• Read on: 08/11/2026 around 19:40 UTC.

Source passages

Pages read on 08/11/2026 around 19:40 UTC. In table excerpts, columns are separated by "/".

Show the 13 sources

Ref.	Quote (original language)	URL
S1↑1↑2↑3↑4↑5↑6+25	"A GLM-5.3-class model that runs on a single NVIDIA B200, or on a Mac Studio with 192 GB or more of unified memory (with mlx-serve)." ; "At 179.9 GB it is 24.6% smaller than the 256-expert UD-IQ2_M (238.6 GB), which does not fit in the memory of one B200 or a 192 GB Mac." ; "every MoE layer keeps 192 of the original 256 routed experts. The model is built from Unsloth's UD-IQ2_M GGUF quantization." ; table: "routed experts per MoE layer / 256 / 192" ; "parameters / 753B / 563B" ; "file size / 238.58 GB (222.19 GiB) / 179.86 GB (167.51 GiB), -24.6%" ; "average bits per weight / (empty) / 2.56" ; "MTP (next-token prediction) layer / yes / removed" ; "HumanEval pass@1 (greedy, thinking off) / not measured / 92.7%" ; "GPQA-Diamond (T=1.0, reasoning_effort=high, 65,536 tokens) / not measured / 84.8%" ; "The 192-experts-per-layer architecture is the one found by the neural architecture search behind GLM-5.3-SLIM-E192. Each MoE layer (3-77) keeps the experts chosen by the search; the other 64 routed experts per layer are removed." ; "The kept experts are copied byte-for-byte, in their original order" ; "No retraining, no recalibration." ; "all other tensors (attention, DSA indexer, shared experts, dense layers 0-2, embeddings, output) are byte-identical to the source" ; "Do not use MTP speculative decoding with this model." ; "behind GLM-5.3-SLIM-E192" ; "the kept router rows (ffn_gate_inp, exp_probs_b) are bit-identical in all 75 MoE layers" ; "ffn_down_exps / 65.9 GiB / IQ3_XXS (71 layers), IQ4_XS (4)" ; "ffn_gate_exps, ffn_up_exps / 2 x 43.7 GiB / IQ2_XXS (74), IQ2_S (1)" ; "attn_q_b, attn_k_b, attn_v_b, attn_kv_a_mqa, indexer / 4.7 GiB / Q8_0" ; "Everything below was run on one NVIDIA B200 (183,359 MiB reported by nvidia-smi)" ; "The weights alone need about 167 GiB of GPU memory" ; "about 100K tokens of KV cache are left for up to 10 parallel requests" ; "generation, tok/s (total), f16 KV / 34.5" (1 sequence) ; "Single-sequence numbers vary by about +-10% between runs" ; "Our run took about 13.5 hours" ; "the client aborted and restarted requests 275 times" ; "At the tested commit the CPU backend cannot run this architecture; the original 256-expert UD-IQ2_M fails the same way." ; "All results below come from our own scripts against llama-server with the settings shown. They are not comparable to the numbers in the official GLM-5.3 card. We did not run the original 256-expert UD-IQ2_M or the full-precision model on the same harness, so the cost of the expert reduction itself is not measured here; for that, see the A/B results of GLM-5.3-SLIM-E192 (FP8) against GLM-5.3 (FP8)." ; "GPQA-Diamond (198 questions) / … / 84.8% (168/198)" ; "HumanEval (164), pass@1 / greedy, thinking off, max 2048 new tokens, one request at a time / 92.7% (152/164)" ; "MMLU, 570-question subset (10 per subject, seed 0) / … / 71.4% (407/570)" ; "answers were run against the official tests in a subprocess" ; "All 1,782 tensors were compared with the source UD-IQ2_M files" ; "Two losses stack: 25% of the routed experts are removed, and the weights are at ~2.6 bits per weight." ; "SLIM's A/B tests of the same expert set (FP8) show a clear loss on Chinese knowledge QA and a smaller loss on science QA; expect the same here." ; "Released under the GLM-5.3 License, the same terms as the base model. Weights by Z.ai; quantization and imatrix by Unsloth; expert selection from GLM-5.3-SLIM-E192"	https://huggingface.co/Scantist/GLM-5.3-SLIM-GGUF
S2↑1↑2↑3↑4	"id":"Scantist/GLM-5.3-SLIM-GGUF" ; "downloads":4938 ; "likes":949 ; "lastModified":"2026-11-03T01:48:48.000Z" ; "createdAt":"2026-11-02T13:11:14.000Z" ; "license:other" ; "license_name":"glm-5.3" ; "gguf":{"total":562687789632 ; "totalFileSize":179861261600	https://huggingface.co/api/models/Scantist/GLM-5.3-SLIM-GGUF
S3↑1↑2↑3↑4↑5	"GLM-5.3 is the most capable open-weights model for coding" ; "Benchmark" table: Terminal Bench 2.1, Terminal Bench 3.0, DeepSWE, NL2Repo, ProgramBench, FrontierSWE, SWE-Marathon, PostTrainBench, CyberGym, ExploitGym, ExploitBench, Toolathlon Verified, AutomationBench, Agents' Last Exam, HLE w/ Tools, GDPval-AA v2 (neither HumanEval nor GPQA-Diamond) ; "license: other" ; "license_name: glm-5.3"	https://huggingface.co/zai-org/GLM-5.3
S4↑1↑2↑3	"Permission is hereby granted, free of charge, to any person or entity […] to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software; to run, deploy, fine-tune, or otherwise modify the Software and create derivative works from it" ; "The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software. The Licensee's use of the Software must comply with applicable laws and regulations." ; "\"Model as a Service\" means giving a third party access to language model inference or fine-tuning (e.g., via API) in a manner that allows such third party to exercise meaningful control over the inputs, parameters, or training data. This does not include (a) end-user products with model capabilities solely embedded within specific features or harnesses, or (b) mere relaying of requests to models hosted by others." ; "If the Licensee or any of its affiliates operates a Model as a Service business, and the aggregate revenue of the Licensee and its affiliates exceeds 10 billion US dollars (or the equivalent in other currencies) in total over any consecutive 12 months, the Licensee must pass Z.AI's security review before using the Software or its derivative works for any commercial purpose." ; "PROVIDED ON AN \"AS IS\" BASIS, WITHOUT WARRANTY OF ANY KIND" ; "Copyright (c) 2026 Z.AI"	https://huggingface.co/zai-org/GLM-5.3/blob/main/LICENSE
S5↑1↑2↑3↑4↑5↑6	"safetensors":{"parameters":{"BF16":2103729152,"F8_E4M3":751226191872,"F32":19456},"total":753329940480} ; "num_experts_per_tok": 8 ; "quant_method": "fp8" ; "createdAt":"2026-08-25T06:42:50.000Z" ; "license:other" ; "downloads":2052042 ; "likes":2397	https://huggingface.co/api/models/zai-org/GLM-5.3
S6↑1↑2↑3↑4	Page response: "Invalid username or password." (HTTP 401, restricted or private repository; not read, not bypassed)	https://huggingface.co/autotrust/GLM-5.3-SLIM
S7↑1↑2	"base_model: zai-org/GLM-5.3" ; "base_model_relation: quantized" ; "See Unsloth Dynamic 3.0 GGUFs for our quantization benchmarks." ; "license_name: glm-5.3"	https://huggingface.co/unsloth/GLM-5.3-GGUF
S8↑1↑2	"We present GPQA, a challenging dataset of 448 multiple-choice questions written by domain experts in biology, physics, and chemistry. We ensure that the questions are high-quality and extremely difficult"	https://arxiv.org/abs/2311.12022
S9↑1↑2	"We propose a new test to measure a text model's multitask accuracy. The test covers 57 tasks including elementary mathematics, US history, computer science, law, and more."	https://arxiv.org/abs/2009.03300
S10↑1↑2	"we introduce speculative decoding - an algorithm to sample from autoregressive models faster without any changes to the outputs"	https://arxiv.org/abs/2211.17192
S11↑	"On HumanEval, a new evaluation set we release to measure functional correctness for synthesizing programs from docstrings"	https://arxiv.org/abs/2107.03374
S12↑	"36GB unified memory" ; "96GB unified memory" ; "256GB unified memory" ; "512GB unified memory" (Mac Studio tech specs)	https://www.apple.com/mac-studio/specs/
R1↑1↑2↑3↑4↑5↑6+2	Our calculation. Z.AI's weights: 751,226,191,872 × 1 byte (FP8) + 2,103,729,152 × 2 bytes (BF16) + 19,456 × 4 bytes (F32) = 755,433,728,000 bytes, about 755 GB S5. Experts removed: 64 × 75 layers (3 to 77) = 4,800 out of 256 × 75 = 19,200, i.e. 25% S1. Size ratio: 755.43 / 179.86 = 4.2 S1S5. Download: 179,861,261,600 bytes × 8 / 10⁹ bit/s = 1,439 s, about 24 min S2. B200: 183,359 MiB × 1,048,576 = 192.3 GB S1. 500-token answer at 34.5 tokens/s = 14.5 s S1. Weights on the B200: 167 GiB × 1.0737 = 179.3 GB, close to 180 GB S1. Check of the 2.56 bits: 179,861,261,600 × 8 / 562,687,789,632 = 2.557 S2.	— (in-house calculation)