← Back to the full write-up on hijoseromero.com

GLM-5.3-Flash on 128 GB: every quant that fits, reasoning low and high

Z.ai's 320B mixture-of-experts (18B active, MIT) on a llama.cpp build of Unsloth's not-yet-merged support branch. Every Unsloth quant in size order until one no longer fits a 128 GB desktop, the same four one-shot tests each time, first at reasoning effort low, then at high. Every output, thinking transcript, partial output and number is linked.

5 of 6files fit93 to 120 GB on disk

120.4 GBbiggest that runspeak 120.5 of 123 GB usable

147.5 GBwhere it breaksfails to load

8.2 to 9.6tok/s at low, every filesize barely changes speed

12/20tests finished at low60 minute cap each

1/20tests finished at highsame files, same cap

The model

ModelGLM-5.3-Flash, by Z.ai

Parameters320B total, 18B active per token (mixture-of-experts)

Architecturehybrid sparse + linear attention, mHC hyper-connections, multi-token prediction head

Context1M tokens trained; these runs used a 65,536-token window

Reasoning levelslow, high, max (default max), plus a clear_thinking switch

LicenseMIT

Runtime herellama.cpp, Unsloth's GLM-5-Next branch build, Vulkan, flash attention off, no MTP

MachineGMKtec EVO-X2, Ryzen AI Max+ 395, 128 GB unified (123 GB usable), CachyOS

model card · Unsloth guide · Unsloth GGUF repo · llama.cpp PR #27754 · first night, smallest file

Unsloth's files, by size

UD-IQ1_S93.1 GB ran

UD-IQ1_M97.6 GB ran

UD-IQ2_XXS101.8 GB ran

UD-Q2_K_XL108.7 GB ran

UD-IQ3_XXS120.4 GB ran

UD-Q3_K_XL147.5 GB does not load

UD-IQ4_XS156.8 GB does not fit

UD-Q4_K_XL199.7 GB does not fit

BF16641.6 GB does not fit

Bars scaled to 200 GB, BF16 runs off the chart. The dashed line is 123 GB usable on this machine. Sizes from the Hugging Face API; the repo has a few more variants between 4-bit and BF16.

Not out of the box: where every dial was set

These numbers are not the defaults. The context window, the reasoning level, flash attention and MTP were all set by hand, and each one changes what the ladder measures.

Dial	This page	Out of the box	Why it matters
Context window	65,536 tokens	1,048,576 (the model card); llama-server's default `-c 0` means "whatever the model says", so out of the box it would try to reserve a 1M-token cache. Unsloth's example command sets no size at all.	set on purpose so the KV cache fits next to a 120 GB file. No test came near it: the DNF partials are about 23k tokens each, so time was the limit, never context.
Reasoning effort	low, then high; bonus leg at max	max	chat-template argument, read at launch. Both ladders ran below the model's default.
clear_thinking	true	false	drops earlier thinking from the conversation; the model card recommends true for chat.
Samplers	temperature 1.0, top_p 0.95, top_k 0, min_p 0, seed 42	temperature 1.0, top_p 0.95	the model card's numbers; seed pinned so every file sees the same dice.
Flash attention	off (`-fa off`)	auto	the branch needs it off for MLA precision (PR #27754). It is also why generation slows as the context fills.
Multi-token prediction (MTP)	off	on in Unsloth Desktop; Unsloth says the llama.cpp branch gained MTP plus faster decoding on Sep 4 ("up to 3.3x faster" at long context)	this build is from Aug 28 and predates that. The binary lists `--spec-type draft-mtp`; not tested here. Every tok/s on this page is the no-MTP number, a rerun on the new branch is the obvious follow-up.
Time budget	60 minutes per test, max_tokens 120,000	none	the cap that produced every DNF.

llama-server -m GLM-5.3-Flash-UD-IQ1_S-00001-of-00003.gguf -c 65536 -ngl 999 -np 1 --jinja -fa off \
  --chat-template-kwargs '{"reasoning_effort": "low", "clear_thinking": true}'   # then "high", then "max"

The short version

Five files fit, 93 to 120 GB. The 147 GB file does not load.
Speed at low: 8.2 to 9.6 tok/s on every file. Size barely moves it.
Tests finished at low: 1-bit S 4/4, 1-bit M 1/4, 2-bit XXS 2/4, 2-bit XL 3/4, 3-bit XXS 2/4.
Tests finished at high: 1 of 20 (1-bit S Ledger, 41.9 min). At max, on the smallest file: 0 of 8, the thinking loops on one word after about 20k tokens.
Every DNF is the 60-minute cap: the model drafts the whole answer inside its thinking, then writes it again.
The Ledger, all six numbers exact: 2-bit XL and 3-bit XXS. Blank page, one token from correct: 1-bit S, 1-bit M, 2-bit XXS.
Signature: "Drawn by GLM" on 1-bit S, 2-bit XL, 3-bit XXS. "Claude" on 2-bit XXS.

The ladder at reasoning low: one machine, the files in size order, same exam

Same prompts, same settings, same seed; only the file gets bigger. "Whole exam" = wall time for the tests that finished; "thinking" = its share of that. Peak memory = total system RAM at the run's highest point, OS included.

Build	File	Peak memory of 123 GB usable	Load	Generation	Finished	Whole exam	Of which thinking	Output tokens
1-bit S, low UD-IQ1_S, 93.1 GB, reasoning low	93.1 GB	96 GB	25 s	9.4 tok/s	4/4	25m 47s	1m 24s	14,266
1-bit M, low UD-IQ1_M, 97.6 GB, reasoning low	97.6 GB	111 GB	30 s	8.9 tok/s	1/4	8m 18s	1m 52s	4,355
2-bit XXS, low UD-IQ2_XXS, 101.8 GB, reasoning low	101.8 GB	103 GB	30 s	9.1 tok/s	2/4	14m 45s	1m 00s	7,924
2-bit XL, low UD-Q2_K_XL, 108.7 GB, reasoning low	108.7 GB	111 GB	30 s	8.7 tok/s	3/4	34m 34s	12m 45s	17,452
3-bit XXS, low UD-IQ3_XXS, 120.4 GB, reasoning low	120.4 GB	120 GB	35 s	8.5 tok/s	2/4	14m 00s	1m 07s	6,999
3-bit XL, low UD-Q3_K_XL, 147.5 GB, bigger than the machine's memory	147.5 GB	-	-	does not fit: llama-server exited during load
Qwen 3.8-Flash-Next 1-bit, medium UD-IQ1_M, 74.5 GB, thinking medium (earlier run, same box, same prompts)	74.5 GB	82 GB	20 s	23.1 tok/s	4/4	25m 21s	14m 08s	34,631
Qwen 3.8-27B 4-bit, medium UD-Q4_K_XL, 17.9 GB, thinking medium (earlier run, same box, same prompts)	17.9 GB	38 GB	9 s	19.9 tok/s	4/4	21m 57s	7m 43s	25,676

rows are the Qwen references from earlier runs (same box, same prompts): they ran with MTP on and flash attention on, so their speed is the yardstick, not a like-for-like build. Every did-not-finish is on the grid below.

The same ladder at reasoning high

Build	File	Peak memory of 123 GB usable	Load	Generation	Finished	Whole exam	Of which thinking	Output tokens
1-bit S, high UD-IQ1_S, 93.1 GB, reasoning high	93.1 GB	95 GB	25 s	7.3 tok/s	1/4	41m 53s	32m 20s	18,286
1-bit M, high UD-IQ1_M, 97.6 GB, reasoning high	97.6 GB	98 GB	30 s	-	0/4	every test hit the 60 minute cap, 4 partial outputs kept (per-test tables below)
2-bit XXS, high UD-IQ2_XXS, 101.8 GB, reasoning high	101.8 GB	102 GB	30 s	-	0/4	every test hit the 60 minute cap, 4 partial outputs kept (per-test tables below)
2-bit XL, high UD-Q2_K_XL, 108.7 GB, reasoning high	108.7 GB	109 GB	30 s	-	0/4	every test hit the 60 minute cap, 4 partial outputs kept (per-test tables below)
3-bit XXS, high UD-IQ3_XXS, 120.4 GB, reasoning high	120.4 GB	119 GB	35 s	-	0/4	every test hit the 60 minute cap, 4 partial outputs kept (per-test tables below)
3-bit XL, high UD-Q3_K_XL, 147.5 GB, bigger than the machine's memory	147.5 GB	-	-	does not fit: llama-server exited during load

Same files, same prompts, same seed, only the reasoning dial moved from low to high. 60-minute cap per test.

1 of 20 finished: 1-bit S Ledger, 41.9 min, 18,286 tokens, 7.3 tok/s, blank page (same let top slip as at low).
19 DNF at 60 min, partial kept for each: 47k to 96k thinking characters, about 23k tokens per hour.
Answer started before the cap on 6 of 19: Ledger on 1-bit M (45k chars), 2-bit XXS (29k), 2-bit XL (23k); Blockfall on 2-bit XL (13k), 3-bit XXS (14k). None runnable.
Eruption and Blind Artist: thinking only, on every file.
Load and peak memory within 2 GB of the low ladder. The 147 GB file fails to load at both settings.

Bonus: reasoning max on the smallest file

GLM-5.3-Flash has three reasoning levels, low, high and max, and max is the model's default. The ladder skipped it because at 9 tok/s nothing would finish inside the hour. This is what happens when you leave the dial where Z.ai set it, on the 93 GB file: first with the ladder's 60-minute cap and 65k context, then with a 3-hour cap per test and a 262k context.

Build	File	Peak memory of 123 GB usable	Load	Generation	Finished	Whole exam	Of which thinking	Output tokens
1-bit S, max UD-IQ1_S, 93.1 GB, reasoning max (the model's default)	93.1 GB	101 GB	25 s	-	0/4	every test hit the 60 minute cap, 4 partial outputs kept (per-test tables below)
1-bit S, max, 3 h cap UD-IQ1_S, 93.1 GB, reasoning max, 180 min per test, 262k context	93.1 GB	105 GB	25 s	-	0/4	every test hit the 180 minute cap, 4 partial outputs kept (per-test tables below)

The 1-bit S at reasoning_effort max, the setting the model ships with. First with the ladder's 60 minute cap and 65k context, then with a 3 hour cap per test and a 262k context (131k if the bigger one fails to load).

60-minute cap, 65k context: 0 of 4 finished. Thinking only, 82k to 93k characters per test, 0 answer characters.
3-hour cap, 262k context: 0 of 4 finished. Thinking only, 224k to 236k characters per test, 0 answer characters. Loaded fine, peak 104.9 GB, no swap, no OOM.
The thinking collapses into one repeated word (Russian "новых") after 70k to 85k characters, about 20k tokens, on every test. The 3-hour run spent its last two hours repeating it.
Same seed, same text: the 3-hour partials are byte-identical to the 60-minute ones up to the point the cap hit. More time and more context changed nothing.
Speed decays with the filling context: about 24k tokens in the first hour, 48k in three.
Verdict for the 93 GB file: max is unusable. Low is the only level that finishes the exam; high finishes the Ledger only.

GLM speed = llama-server's generation timing. Machine and runtime details are in the model card at the top.

How this ran before llama.cpp supports it

The problem, still

GLM-5.3-Flash uses a brand-new architecture (glm5next: hybrid linear + sparse attention, 288 routed experts, "hyper-connections"). Stock llama.cpp master still refuses the file: unknown model architecture: 'glm5next'. The working support is Unsloth's llama.cpp PR #27754, "model: add GLM-5-Next (GLM-5.3-Flash)", branch glm5next/upstream on Unsloth's fork. It requires flash attention off (-fa off), and MTP is in the file but the branch does not use it yet. Everything here ran on Vulkan (RADV) on a Strix Halo desktop.

The recipe

Same box as the Qwen runs: GMKtec EVO-X2, Ryzen AI Max+ 395, 128 GB unified memory (123 GB usable), Radeon 8060S, CachyOS. Branch build in its own folder; stock master untouched.

1. Build the branch with Vulkan.

git clone --branch glm5next/upstream --depth 1 https://github.com/unslothai/llama.cpp llama.cpp-glm5next
cd llama.cpp-glm5next
cmake -B build -DGGML_VULKAN=ON -DCMAKE_BUILD_TYPE=Release
cmake --build build --config Release -j $(nproc) --target llama-server

2. Download the quants in size order. From unsloth/GLM-5.3-Flash-GGUF. The full list, smallest first: UD-IQ1_S 93.1 GB, UD-IQ1_M 97.6, UD-IQ2_XXS 101.8, UD-Q2_K_XL 108.7, UD-IQ3_XXS 120.4, UD-Q3_K_XL 147.5, UD-IQ4_XS 156.8, UD-Q4_K_XL 199.7, then Q5, Q6, Q8 and BF16 (642 GB). Unsloth's own table says 1-bit needs 100 GB, 2-bit 115, 3-bit 128 to 150. With 123 GB usable the ladder was: every file in size order through the 3-bit XL, and stop at the first one that does not fit. The Q3_K_XL (147.5 GB) is bigger than the machine's memory, so it cannot be resident; it went through the same gate anyway so the break is measured, not assumed.

hf download unsloth/GLM-5.3-Flash-GGUF --include "UD-IQ1_S/*"
hf download unsloth/GLM-5.3-Flash-GGUF --include "UD-IQ1_M/*"
hf download unsloth/GLM-5.3-Flash-GGUF --include "UD-IQ2_XXS/*"
hf download unsloth/GLM-5.3-Flash-GGUF --include "UD-Q2_K_XL/*"
hf download unsloth/GLM-5.3-Flash-GGUF --include "UD-IQ3_XXS/*"
hf download unsloth/GLM-5.3-Flash-GGUF --include "UD-Q3_K_XL/*"

3. One server per file, the same exam. 65k context, flash attention off, reasoning effort set at launch (the branch reads it from the chat template), model-card samplers: temperature 1.0, top_p 0.95. Four prompts, one shot each, seed 42, 60 minutes per test, then the server is stopped and the next file loads. The runner records wall time, generation and prefill speed from llama-server's own timings, output tokens, thinking share, peak system memory, load time.

build/bin/llama-server -m <first shard>.gguf --host 127.0.0.1 --port 8899 -np 1 -c 65536 -ngl 999 \
  --jinja --no-webui -fa off --chat-template-kwargs '{"reasoning_effort":"low","clear_thinking":true}'

The high ladder is the same command with "reasoning_effort":"high".

4. The does-not-fit gate. A file that is larger than free memory does not fail cleanly on this box: llama.cpp maps it from disk and the model "runs" at a token every few seconds while the kernel pages weights in and out. To keep that from eating four hours of time limits, the runner sends an 8-token warmup ping first; if it does not come back within 5 minutes the leg is recorded as does-not-fit and the next file loads. Legs that hit that gate are listed under Errors with the exact message.

What the server log says about Vulkan. On every layer the branch prints fused DeepSeek V4 HC pre/comb/post not supported, set to disabled: the hyper-connection ops have no fused Vulkan kernel yet and take the slow path. The speeds on this page are a first-week floor for this backend, not the model's ceiling.

Test results

Per test: a mid-play frame of every build that finished (click to open the file the model wrote), and the numbers.

BLOCKFALL a complete falling-blocks game, one prompt

Browser check: two finishers at reasoning low (1-bit S, 2-bit XL), both reach game over with zero console errors; the other three files hit the hour while still drafting inside their thinking

1-bit S, low BLOCKFALL — **1-bit S, low**: game over reached; score 0; ghost piece; 0 console errors

2-bit XL, low BLOCKFALL — **2-bit XL, low**: game over reached; score 0; 0 console errors

Build	Total time	Thinking	Output	Gen speed	HTML size	Files
1-bit S, low	5m 22s	- (0%)	3,037	9.6 tok/s	8 KB	output · raw
1-bit M, low	DNF 60 min cap	-	-	-	-	no partial kept
2-bit XXS, low	DNF 60 min cap	-	-	-	-	no partial kept
2-bit XL, low	5m 37s	0m 01s (0%)	3,074	9.2 tok/s	8 KB	output · thinking · raw
3-bit XXS, low	DNF 60 min cap	77,180 chars	0 answer chars	-	-	partial output
3-bit XL, low	did not fit bigger than the machine's memory, llama-server exited during load
1-bit S, high	DNF 60 min cap	84,418 chars	0 answer chars	-	-	partial output
1-bit M, high	DNF 60 min cap	81,974 chars	0 answer chars	-	-	partial output
2-bit XXS, high	DNF 60 min cap	82,877 chars	0 answer chars	-	-	partial output
2-bit XL, high	DNF 60 min cap	69,095 chars	13,393 answer chars	-	-	partial output
3-bit XXS, high	DNF 60 min cap	63,743 chars	14,460 answer chars	-	-	partial output
3-bit XL, high	did not fit bigger than the machine's memory, llama-server exited during load
1-bit S, max	DNF 60 min cap	86,603 chars	0 answer chars	-	-	partial output
1-bit S, max, 3 h cap	DNF 180 min cap	233,711 chars	0 answer chars	-	-	partial output

Two finishers at low: the 1-bit S (5.4 min, 3,037 tokens: five keys, ghost piece, next preview, game over, zero errors) and the 2-bit XL (5.6 min, 3,074 tokens, same minus the ghost piece). The 1-bit M, 2-bit XXS and 3-bit XXS all hit the hour. The 3-bit XXS partial shows why: a full requirements checklist, then the complete game (CSS, canvas, lock delay, sound) written inside the thinking, no answer started. The two files that finished skipped that step (0% thinking share) and wrote the game once.

The exact prompt (one shot, sent as-is)

Create a complete falling-blocks puzzle game in a single self-contained HTML file. No external libraries, no internet resources, no images: everything inline (HTML, CSS, JavaScript, canvas rendering).

Requirements:

- 10 wide by 20 tall playfield rendered on a canvas.
- Seven distinct piece shapes, each made of four squares, each with its own color.
- Controls: left/right arrows move, up arrow rotates, down arrow soft-drops, spacebar hard-drops.
- Rotation must respect walls and stacked blocks (no clipping through anything).
- Completed horizontal lines clear, rows above fall down, and clearing multiple lines at once scores more.
- Score, lines cleared, and level displayed. Speed increases with level.
- Next-piece preview box.
- Game over when the stack reaches the top, with a visible game-over state and a restart key.
- Clean, readable dark visual style.

Output only the complete HTML file.

open as plain text ↗

ERUPTION a volcano physics simulation with live controls

Browser check: one finisher at reasoning low (1-bit S: 900 particles at 10 s, zero errors); every other file hit the hour, the 3-bit XXS 12,000 characters into its answer

1-bit S, low ERUPTION — **1-bit S, low**: 900 particles on screen at 10 s; 50 fps; 0 console errors

Build	Total time	Thinking	Output	Gen speed	HTML size	Files
1-bit S, low	8m 04s	0m 04s (0%)	4,426	9.2 tok/s	10 KB	output · thinking · raw
1-bit M, low	DNF 60 min cap	-	-	-	-	no partial kept
2-bit XXS, low	DNF 60 min cap	-	-	-	-	no partial kept
2-bit XL, low	DNF 60 min cap	-	-	-	-	no partial kept
3-bit XXS, low	DNF 60 min cap	62,462 chars	12,045 answer chars	-	-	partial output
3-bit XL, low	did not fit bigger than the machine's memory, llama-server exited during load
1-bit S, high	DNF 60 min cap	81,274 chars	0 answer chars	-	-	partial output
1-bit M, high	DNF 60 min cap	95,539 chars	0 answer chars	-	-	partial output
2-bit XXS, high	DNF 60 min cap	82,201 chars	0 answer chars	-	-	partial output
2-bit XL, high	DNF 60 min cap	85,787 chars	0 answer chars	-	-	partial output
3-bit XXS, high	DNF 60 min cap	78,082 chars	0 answer chars	-	-	partial output
3-bit XL, high	did not fit bigger than the machine's memory, llama-server exited during load
1-bit S, max	DNF 60 min cap	92,648 chars	0 answer chars	-	-	partial output
1-bit S, max, 3 h cap	DNF 180 min cap	235,760 chars	0 answer chars	-	-	partial output

One finisher at low: the 1-bit S (8.1 min, 900 particles at 10 s, live controls, zero errors). Every other file ran into the hour; the 2-bit XL did it twice, at 240 minutes on its first night and at 60 on the second. The 3-bit XXS partial: 62,000 characters of thinking, then 12,000 characters of a real answer cut off mid-file. Eruption is the test this model drafts longest on.

The exact prompt (one shot, sent as-is)

Create a 2D physics simulation of a volcano erupting, in a single self-contained HTML file. No external libraries: everything inline, rendered on a canvas.

Requirements:

- Side view: a volcano cone with a crater, ground, and sky.
- A pressure meter that builds over time; when it maxes out, the volcano erupts, then pressure resets and builds again on its own.
- Eruptions launch glowing lava particles from the crater with varied velocities. Particles follow gravity in a realistic arc, land on the slopes, and slide or settle downhill.
- Lava cools over its lifetime: bright yellow-orange when launched, fading to dark rock, then disappearing.
- An ash cloud of smoke particles that rises from the crater and drifts with the wind.
- Three sliders the user can move while it runs: eruption power, wind strength and direction, and particle amount.
- A live on-screen counter showing active particle count and frames per second. The simulation must stay smooth with at least 1000 active particles.

Output only the complete HTML file.

open as plain text ↗

THE LEDGER clean a messy CSV and build a correct dashboard

Browser check: 2-bit XL and 3-bit XXS render all six key numbers exact with zero errors; 1-bit S, 1-bit M and 2-bit XXS render blank because of a single-token script error each (numbers correct in the source; hand-fixed renders shown)

1-bit S, low THE LEDGER — **1-bit S, low**: only 0/6 key numbers found (check by hand); duplicate id not mentioned; 1 console/page error(s): page: Identifier 'top' has already been declared

1-bit S, low THE LEDGER, fixed by hand — **1-bit S, low, fixed by hand**: one identifier renamed (`top` to `topRegion`, it collided with the browser's built-in `window.top`); all 6 key numbers exact, 0 console errors. This is what an agent harness would have shipped after reading the console.

1-bit M, low THE LEDGER — **1-bit M, low**: only 0/6 key numbers found (check by hand); duplicate id not mentioned; 1 console/page error(s): page: Identifier 'top' has already been declared

2-bit XXS, low THE LEDGER — **2-bit XXS, low**: only 0/6 key numbers found (check by hand); duplicate id not mentioned; 1 console/page error(s): page: Unexpected token '.'

2-bit XXS, low THE LEDGER, fixed by hand — **2-bit XXS, low, fixed by hand**: one entry deleted from an object literal (`'texas '.trim().toLowerCase():'Texas'`, an expression is not a valid key without brackets, so the whole script failed to parse); all 6 key numbers exact, duplicate flagged, 0 console errors. This is what an agent harness would have shipped after reading the console.

2-bit XL, low THE LEDGER — **2-bit XL, low**: all 6 key numbers exact; 0 console errors

3-bit XXS, low THE LEDGER — **3-bit XXS, low**: all 6 key numbers exact; duplicate id not mentioned; 0 console errors

1-bit S, high THE LEDGER — **1-bit S, high**: only 0/6 key numbers found (check by hand); duplicate id not mentioned; 1 console/page error(s): page: Identifier 'top' has already been declared

1-bit S, high THE LEDGER, fixed by hand — **1-bit S, high, fixed by hand**: the same one-token slip as the low run, 18,000 tokens later: `top` renamed to `topRegion` (two lines); all 6 key numbers exact, 0 console errors. This is what an agent harness would have shipped after reading the console.

Build	Total time	Thinking	Output	Gen speed	HTML size	Files
1-bit S, low	5m 20s	0m 17s (5%)	2,894	9.4 tok/s	7 KB	output · thinking · raw
1-bit M, low	8m 18s	1m 52s (22%)	4,355	8.9 tok/s	8 KB	output · thinking · raw
2-bit XXS, low	7m 21s	0m 55s (12%)	3,863	9.0 tok/s	8 KB	output · thinking · raw
2-bit XL, low	10m 11s	0m 53s (8%)	5,193	8.6 tok/s	14 KB	output · thinking · raw
3-bit XXS, low	7m 56s	0m 52s (11%)	3,884	8.3 tok/s	8 KB	output · thinking · raw
3-bit XL, low	did not fit bigger than the machine's memory, llama-server exited during load
1-bit S, high	41m 53s	32m 20s (77%)	18,286	7.3 tok/s	12 KB	output · thinking · raw
1-bit M, high	DNF 60 min cap	46,741 chars	45,159 answer chars	-	-	partial output
2-bit XXS, high	DNF 60 min cap	57,722 chars	28,979 answer chars	-	-	partial output
2-bit XL, high	DNF 60 min cap	55,959 chars	23,491 answer chars	-	-	partial output
3-bit XXS, high	DNF 60 min cap	88,665 chars	0 answer chars	-	-	partial output
3-bit XL, high	did not fit bigger than the machine's memory, llama-server exited during load
1-bit S, max	DNF 60 min cap	88,267 chars	0 answer chars	-	-	partial output
1-bit S, max, 3 h cap	DNF 180 min cap	230,161 chars	0 answer chars	-	-	partial output

The one test with a provable answer, and the one that separates the files. 2-bit XL (10.2 min) and 3-bit XXS (7.9 min): all six numbers exact (California 16,091.26, Ohio 6,498.24, Texas 9,493.50, quarter totals 16,079.59 and 16,003.41, grand total 32,083.00), refunds negative, duplicate dropped, headline that reads like a person wrote it, zero console errors. 1-bit S and 1-bit M: blank page, let top at global scope collides with the browser's built-in window.top; the numbers are right in the source. 2-bit XXS: blank page, the region map has 'texas '.trim().toLowerCase() as an object key, which is not valid JavaScript without brackets, so nothing runs. Both fixes are one token; the fixed renders are next to the broken frames, every number exact. A one-shot prompt does not get that second look; an agent harness does.

The exact prompt (one shot, sent as-is)

Below is a messy sales CSV export. Build a single self-contained HTML dashboard file (no external libraries, everything inline) that cleans this data and reports on it.

Cleaning rules:

- order_id must be unique: if an id appears twice, keep the first row and drop the rest.
- Rows whose amount is not a number (like N/A or unknown) are excluded from revenue, but report how many you excluded.
- Negative amounts are refunds and must count as negative revenue.
- Region names must be normalized (TX, tx, Texas, texas are all Texas; CA is California; OH is Ohio).
- Dates appear in multiple formats; parse them all. Q1 = Jan-Mar, Q2 = Apr-Jun.

The dashboard must show:

1. A pivot table of total revenue by region (rows) and quarter (columns), with row totals, column totals, and a grand total, formatted as dollars with two decimals.
2. A bar chart of revenue by region drawn with inline SVG or canvas (no chart libraries).
3. A data quality panel: rows received, duplicates dropped, rows excluded, rows counted.
4. A one-line headline stating the top region and its total.

The numbers displayed must be computed correctly from the data. Output only the complete HTML file.

```
order_id,order_date,region,product,amount_usd
1001,2025-01-06,Texas,widget,"1,240.00"
1002,1/9/2025,CA,gadget,$980
1003,2025-01-14,TX,gizmo,312.50
1004,"Jan 21, 2025",california,widget,"1,764.00"
1005,2025-01-27,Ohio,gadget,455
1006,2/3/2025,texas ,widget,N/A
1007,2025-02-10,OH,gizmo,"1,102.25"
1008,"Feb 14, 2025",California,widget,"2,310.00"
1009,2025-02-18,TX,gadget,-450.00
1010,2025-02-24,Texas,gizmo,870.10
1011,3/3/2025,Ohio,widget,"$1,999.99"
1012,2025-03-08,CA,gizmo,640.00
1008,"Feb 14, 2025",California,widget,"2,310.00"
1013,"Mar 12, 2025",tx,gadget,"1,580.00"
1014,2025-03-19,ohio,gizmo,225.75
1015,2025-03-27,California,gadget,"3,050.00"
1016,4/2/2025,TX,widget,990.00
1017,2025-04-07,Ohio,gadget,unknown
1018,"Apr 11, 2025",CA,widget,"1,420.50"
1019,2025-04-16,Texas,gizmo,"2,075.00"
1020,2025-04-22,OH,widget,$760.25
1021,5/1/2025,California,gizmo,"1,111.11"
1022,2025-05-06,TX,gadget,845.00
1023,"May 13, 2025",Ohio,gizmo,-220.00
1024,2025-05-19,texas,widget,"1,300.00"
1025,2025-05-28,CA,gadget,"2,940.40"
1026,6/4/2025,Ohio,widget,515.00
1027,2025-06-10,California,widget,"1,875.25"
1028,"Jun 17, 2025",TX,gizmo,N/A
1029,2025-06-23,OH,gadget,"1,660.00"
1030,2025-06-30,Texas,gadget,730.90
```

open as plain text ↗

BLIND ARTIST draw a scene in pure SVG, no eyes allowed

Browser check: four valid SVGs, zero errors, servers on the forks and a dog in a hard hat every time; signed 'drawn by GLM' three times and 'Claude' once (2-bit XXS)

1-bit S, low BLIND ARTIST — **1-bit S, low**: signed "drawn by GLM"; valid SVG; 0 console errors

2-bit XXS, low BLIND ARTIST — **2-bit XXS, low**: signed "Claude"; valid SVG; 0 console errors

2-bit XL, low BLIND ARTIST — **2-bit XL, low**: signed "drawn by GLM"; valid SVG; 0 console errors

3-bit XXS, low BLIND ARTIST — **3-bit XXS, low**: signed "Drawn by GLM"; valid SVG; 0 console errors

Build	Total time	Thinking	Output	Gen speed	HTML size	Files
1-bit S, low	7m 01s	1m 03s (15%)	3,909	9.4 tok/s	8 KB	output · thinking · raw
1-bit M, low	DNF 60 min cap	-	-	-	-	no partial kept
2-bit XXS, low	7m 24s	0m 05s (1%)	4,061	9.2 tok/s	9 KB	output · thinking · raw
2-bit XL, low	18m 46s	11m 51s (63%)	9,185	8.2 tok/s	7 KB	output · thinking · raw
3-bit XXS, low	6m 05s	0m 15s (4%)	3,115	8.6 tok/s	7 KB	output · thinking · raw
3-bit XL, low	did not fit bigger than the machine's memory, llama-server exited during load
1-bit S, high	DNF 60 min cap	75,546 chars	0 answer chars	-	-	partial output
1-bit M, high	DNF 60 min cap	88,586 chars	0 answer chars	-	-	partial output
2-bit XXS, high	DNF 60 min cap	84,936 chars	0 answer chars	-	-	partial output
2-bit XL, high	DNF 60 min cap	71,042 chars	0 answer chars	-	-	partial output
3-bit XXS, high	DNF 60 min cap	84,663 chars	0 answer chars	-	-	partial output
3-bit XL, high	did not fit bigger than the machine's memory, llama-server exited during load
1-bit S, max	DNF 60 min cap	81,839 chars	0 answer chars	-	-	partial output
1-bit S, max, 3 h cap	DNF 180 min cap	224,279 chars	0 answer chars	-	-	partial output

Four finished drawings, all valid SVG, zero errors, all with the servers on the forks, a dog driving in a hard hat, sun, clouds, house. 1-bit S (7.0 min): a ladder, a cartoon dog with goggles, "drawn by GLM". 2-bit XXS (7.4 min): the cleanest forklift of the set, three servers with LEDs, a stacked ladder, a smiling sun, signed "Claude". 2-bit XL (18.8 min, 63% of it thinking): flat illustration style, orange forklift, "drawn by GLM". 3-bit XXS (6.1 min): tall mast, three servers, a bear-like dog, "Drawn by GLM". The 1-bit M ran into the hour.

The exact prompt (one shot, sent as-is)

Using only raw inline SVG code (no JavaScript, no external images, no libraries), draw this exact scene as a single self-contained HTML file:

A happy dog wearing a yellow hard hat is driving a forklift. The forklift is lifting a rack of three stacked computer servers on its front forks. The servers have small colored indicator lights. The scene is outdoors: flat ground, a sun in the top left, two clouds, and a small house far in the background on the right. The forklift has two wheels touching the ground and the dog sits in the driver seat holding the steering wheel.

Layering must be correct (wheels touch the ground, the dog is inside the forklift, the hat is on the dog's head, the servers sit on the forks). Sign the artwork by writing your own model name in small text in the bottom right corner.

open as plain text ↗

How the browser check was done

Each finished output opened in headless Chromium, console captured. Blockfall: all keys pressed, then 60 hard drops. Eruption: 10.5 s wait, counters read. Ledger: the six key numbers searched in the rendered page. Blind Artist: SVG validity and signature text. The screenshots on this page are those frames. Did-not-finish legs have no output file, so they have no screenshot; from the 3-bit legs on, the text generated before the cap is linked as "partial output".

Every run on one grid: finished, did not finish, did not fit

Build	Blockfall	Eruption	The Ledger	Blind Artist	Finished
Reasoning low
1-bit S, low 93.1 GB	5m 22s9.6 tok/s · 3,037 tok · output	8m 04s9.2 tok/s · 4,426 tok · output	5m 20s9.4 tok/s · 2,894 tok · output	7m 01s9.4 tok/s · 3,909 tok · output	4/4
1-bit M, low 97.6 GB	DNFno partial	DNFno partial	8m 18s8.9 tok/s · 4,355 tok · output	DNFno partial	1/4
2-bit XXS, low 101.8 GB	DNFno partial	DNFno partial	7m 21s9.0 tok/s · 3,863 tok · output	7m 24s9.2 tok/s · 4,061 tok · output	2/4
2-bit XL, low 108.7 GB	5m 37s9.2 tok/s · 3,074 tok · output	DNFno partial	10m 11s8.6 tok/s · 5,193 tok · output	18m 46s8.2 tok/s · 9,185 tok · output	3/4
3-bit XXS, low 120.4 GB	DNF77k thinking · 0k answer · partial	DNF62k thinking · 12k answer · partial	7m 56s8.3 tok/s · 3,884 tok · output	6m 05s8.6 tok/s · 3,115 tok · output	2/4
3-bit XL, low 147.5 GB	did not fitbigger than the machine's memory, llama-server exited during load
Reasoning high
1-bit S, high 93.1 GB	DNF84k thinking · 0k answer · partial	DNF81k thinking · 0k answer · partial	41m 53s7.3 tok/s · 18,286 tok · output	DNF75k thinking · 0k answer · partial	1/4
1-bit M, high 97.6 GB	DNF81k thinking · 0k answer · partial	DNF95k thinking · 0k answer · partial	DNF46k thinking · 45k answer · partial	DNF88k thinking · 0k answer · partial	0/4
2-bit XXS, high 101.8 GB	DNF82k thinking · 0k answer · partial	DNF82k thinking · 0k answer · partial	DNF57k thinking · 28k answer · partial	DNF84k thinking · 0k answer · partial	0/4
2-bit XL, high 108.7 GB	DNF69k thinking · 13k answer · partial	DNF85k thinking · 0k answer · partial	DNF55k thinking · 23k answer · partial	DNF71k thinking · 0k answer · partial	0/4
3-bit XXS, high 120.4 GB	DNF63k thinking · 14k answer · partial	DNF78k thinking · 0k answer · partial	DNF88k thinking · 0k answer · partial	DNF84k thinking · 0k answer · partial	0/4
3-bit XL, high 147.5 GB	did not fitbigger than the machine's memory, llama-server exited during load
Bonus: reasoning max, smallest file
1-bit S, max 93.1 GB	DNF86k thinking · 0k answer · partial	DNF92k thinking · 0k answer · partial	DNF88k thinking · 0k answer · partial	DNF81k thinking · 0k answer · partial	0/4
1-bit S, max, 3 h cap 93.1 GB	DNF233k thinking · 0k answer · partial	DNF235k thinking · 0k answer · partial	DNF230k thinking · 0k answer · partial	DNF224k thinking · 0k answer · partial	0/4

Build

Blockfall

Eruption

The Ledger

Blind Artist

Finished

Reasoning low

1-bit S, low
93.1 GB

5m 22s9.6 tok/s · 3,037 tok · output

8m 04s9.2 tok/s · 4,426 tok · output

5m 20s9.4 tok/s · 2,894 tok · output

7m 01s9.4 tok/s · 3,909 tok · output

4/4

1-bit M, low
97.6 GB

DNFno partial

8m 18s8.9 tok/s · 4,355 tok · output

DNFno partial

1/4

2-bit XXS, low
101.8 GB

DNFno partial

7m 21s9.0 tok/s · 3,863 tok · output

7m 24s9.2 tok/s · 4,061 tok · output

2/4

2-bit XL, low
108.7 GB

5m 37s9.2 tok/s · 3,074 tok · output

DNFno partial

10m 11s8.6 tok/s · 5,193 tok · output

18m 46s8.2 tok/s · 9,185 tok · output

3/4

3-bit XXS, low
120.4 GB

DNF77k thinking · 0k answer · partial

DNF62k thinking · 12k answer · partial

7m 56s8.3 tok/s · 3,884 tok · output

6m 05s8.6 tok/s · 3,115 tok · output

2/4

3-bit XL, low
147.5 GB

did not fitbigger than the machine's memory, llama-server exited during load

Reasoning high

1-bit S, high
93.1 GB

DNF84k thinking · 0k answer · partial

DNF81k thinking · 0k answer · partial

41m 53s7.3 tok/s · 18,286 tok · output

DNF75k thinking · 0k answer · partial

1/4

1-bit M, high
97.6 GB

DNF81k thinking · 0k answer · partial

DNF95k thinking · 0k answer · partial

DNF46k thinking · 45k answer · partial

DNF88k thinking · 0k answer · partial

0/4

2-bit XXS, high
101.8 GB

DNF82k thinking · 0k answer · partial

DNF57k thinking · 28k answer · partial

DNF84k thinking · 0k answer · partial

0/4

2-bit XL, high
108.7 GB

DNF69k thinking · 13k answer · partial

DNF85k thinking · 0k answer · partial

DNF55k thinking · 23k answer · partial

DNF71k thinking · 0k answer · partial

0/4

3-bit XXS, high
120.4 GB

DNF63k thinking · 14k answer · partial

DNF78k thinking · 0k answer · partial

DNF88k thinking · 0k answer · partial

DNF84k thinking · 0k answer · partial

0/4

3-bit XL, high
147.5 GB

did not fitbigger than the machine's memory, llama-server exited during load

Bonus: reasoning max, smallest file

1-bit S, max
93.1 GB

DNF86k thinking · 0k answer · partial

DNF92k thinking · 0k answer · partial

DNF88k thinking · 0k answer · partial

DNF81k thinking · 0k answer · partial

0/4

1-bit S, max, 3 h cap
93.1 GB

DNF233k thinking · 0k answer · partial

DNF235k thinking · 0k answer · partial

DNF230k thinking · 0k answer · partial

DNF224k thinking · 0k answer · partial

0/4

time finished, wall time · DNF hit the time cap (60 minutes on the ladder, 180 on the 3-hour bonus rows), the text written so far is the partial (thinking and answer characters) · did not fit llama-server could not load the file. Every cell links the model's raw file.

Raw model outputs

Per build: the numbers file, and for each test the output the model wrote, its thinking transcript, the raw response, and the partial where the cap hit.

Build	Files
1-bit S, low	results.json · 01-blockfall.html · 01-blockfall.raw.txt · 03-eruption.html · 03-eruption.thinking.txt · 03-eruption.raw.txt · 04-the-ledger.html · 04-the-ledger.thinking.txt · 04-the-ledger.raw.txt · 05-blind-artist.html · 05-blind-artist.thinking.txt · 05-blind-artist.raw.txt
1-bit M, low	results.json · 04-the-ledger.html · 04-the-ledger.thinking.txt · 04-the-ledger.raw.txt
2-bit XXS, low	results.json · 04-the-ledger.html · 04-the-ledger.thinking.txt · 04-the-ledger.raw.txt · 05-blind-artist.html · 05-blind-artist.thinking.txt · 05-blind-artist.raw.txt
2-bit XL, low	results.json · 01-blockfall.html · 01-blockfall.thinking.txt · 01-blockfall.raw.txt · 04-the-ledger.html · 04-the-ledger.thinking.txt · 04-the-ledger.raw.txt · 05-blind-artist.html · 05-blind-artist.thinking.txt · 05-blind-artist.raw.txt
3-bit XXS, low	results.json · 01-blockfall.partial.txt · 03-eruption.partial.txt · 04-the-ledger.html · 04-the-ledger.thinking.txt · 04-the-ledger.raw.txt · 05-blind-artist.html · 05-blind-artist.thinking.txt · 05-blind-artist.raw.txt
3-bit XL, low	results.json
1-bit S, high	results.json · 01-blockfall.partial.txt · 03-eruption.partial.txt · 04-the-ledger.html · 04-the-ledger.thinking.txt · 04-the-ledger.raw.txt · 05-blind-artist.partial.txt
1-bit M, high	results.json · 01-blockfall.partial.txt · 03-eruption.partial.txt · 04-the-ledger.partial.txt · 05-blind-artist.partial.txt
2-bit XXS, high	results.json · 01-blockfall.partial.txt · 03-eruption.partial.txt · 04-the-ledger.partial.txt · 05-blind-artist.partial.txt
2-bit XL, high	results.json · 01-blockfall.partial.txt · 03-eruption.partial.txt · 04-the-ledger.partial.txt · 05-blind-artist.partial.txt
3-bit XXS, high	results.json · 01-blockfall.partial.txt · 03-eruption.partial.txt · 04-the-ledger.partial.txt · 05-blind-artist.partial.txt
3-bit XL, high	results.json
1-bit S, max	results.json · 01-blockfall.partial.txt · 03-eruption.partial.txt · 04-the-ledger.partial.txt · 05-blind-artist.partial.txt
1-bit S, max, 3 h cap	results.json · 01-blockfall.partial.txt · 03-eruption.partial.txt · 04-the-ledger.partial.txt · 05-blind-artist.partial.txt

Sources

Claim on this page	Direct source
Model facts (320B total, 18B active, hybrid sparse + linear attention, mHC, 1M context, MIT), samplers, reasoning_effort levels	GLM-5.3-Flash model card
The files that ran, their sizes, Unsloth's memory table	Unsloth guide · the GGUF repo
Why a branch build was needed, flash attention off, MTP unused	llama.cpp PR #27754, "model: add GLM-5-Next (GLM-5.3-Flash)"
The first night on the smallest file (the page this one extends)	GLM-5.3-Flash, the smallest quant
The Qwen reference rows (same tests, same box)	Qwen 3.8-Flash-Next results page · the thinking-dial results page

These are affiliate links. If you buy through them I may earn a small commission at no extra cost to you. As an Amazon Associate I earn from qualifying purchases.