GLM-5.3-Flash: which quant works on a Strix Halo

Last time I got the smallest GLM-5.3-Flash file to run on my 128 GB Strix Halo desktop and asked whether you should bother. This time I wanted the full answer: Unsloth ships this model in twelve files from 93 GB to 642 GB, so I downloaded every one that had a chance of fitting, loaded them in size order, ran the same four one-shot tests on each at reasoning low, then did the whole ladder again at high, and kept going until a file refused to load. Then, because the model's default setting is max, I gave max a shot on the smallest file, first with the usual 60-minute cap, then with a 3-hour cap and a 262k context.

The short version: five files fit, speed does not move with size, the smallest file is the only one that finishes the exam, and the two higher reasoning levels do not work on this box. Here is what that looked like.

The ladder

Same box as every run on this channel: GMKtec EVO-X2, Ryzen AI Max+ 395, 128 GB unified memory with 123 GB usable, CachyOS. Same llama.cpp branch as the first night, because stock master still refuses the file: the support lives on Unsloth's pull request #27754, which is still open, and that branch needs flash attention off and does not use MTP. Vulkan, 65k context, seed 42, one shot per test, 60 minutes per test, no harness, no tool calls.

The files, smallest first: 1-bit S at 93 GB, 1-bit M at 98, 2-bit XXS at 102, 2-bit XL at 109, 3-bit XXS at 120, then the 3-bit XL at 147 GB. Unsloth's own table says 1-bit needs about 100 GB, 2-bit 115, 3-bit 128 to 150, so on paper the ladder should have stopped at 2-bit. It did not. Every file up to 120 GB loaded in 25 to 35 seconds, and the 3-bit XXS peaked at 120.5 GB with 5 GB of headroom and never touched swap. The 147 GB file went through the same gate so the break would be measured instead of assumed, and it did not load at either setting.

One warning if you try this on your daily driver: at the 3-bit XXS, anything else running in the background was enough to make the leg fail. This is not a dedicated inference box, it is my PC, and for these runs it had to behave like one.

Reasoning low: size barely matters, the hour does

Speed is flat. 8.2 to 9.6 tokens a second on every file and every test. The 120 GB file is about 10% slower than the 93 GB file. That is the memory bandwidth of this box talking, not the model. The 1-bit Qwen 3.8-Flash-Next on the same machine does 22 to 24, more than double.
Tests finished, by file: 1-bit S 4 of 4, 1-bit M 1 of 4, 2-bit XXS 2 of 4, 2-bit XL 3 of 4, 3-bit XXS 2 of 4. Ten of the twenty tests hit the 60-minute cap.
The partials explain the cap. From the 3-bit leg on, the runner keeps whatever the model wrote before the hour ran out. It is not a loop. At low, the model plans and drafts the whole file inside its thinking, then starts over to write the real answer. Two drafts at 9 tokens a second do not fit in an hour. The 3-bit XXS Blockfall partial is the complete game, CSS, canvas, lock delay and sound, written inside the thinking with the answer never started.

Blockfall finished on the 1-bit S and the 2-bit XL, both in about five and a half minutes with zero console errors; the 2-bit version is the more polished of the two and stops its pieces where they should stop. Eruption finished on the 1-bit S only, the same single explosion and volcano-that-is-not-a-volcano from the first video, and every other file ran into the hour on it. Eruption is the test this model drafts longest on.

The Ledger: where size finally shows

The Ledger is the one test with a provable answer: clean a messy CSV and build a dashboard with six numbers that are either right or wrong. The 2-bit XL and the 3-bit XXS get all six exact (California 16,091.26, Ohio 6,498.24, Texas 9,493.50, the two quarter totals, a 32,083.00 grand total), refunds negative, duplicate dropped, zero console errors, and a headline that reads like a person wrote it. This is the first time on this box that the model looked like the GLM people show off in their presentation demos.

The 1-bit S and 1-bit M ship a blank page: the script declares a variable called top at global scope, which collides with the browser's built-in window.top, so nothing draws. The numbers in the source are right. The 2-bit XXS ships a different blank page: an object key written as an expression without brackets, which is not valid JavaScript. Both fixes are one token. I put the hand-fixed renders next to the broken frames on the results page, every number exact, because that is what an agent harness would have shipped: it reads the console and fixes the crash before it hands you the file. A one-shot prompt does not get that second look.

Blind Artist: it signs GLM, mostly

Four finished drawings, all valid SVG, zero errors, all with the servers on the forks and a dog driving in a hard hat. The 1-bit S signs "drawn by GLM". The 2-bit XL signs "drawn by GLM". The 3-bit XXS signs "Drawn by GLM". The 2-bit XXS, which drew the cleanest forklift of the set, signed "Claude", and its thinking transcript says so out loud: carefully layered, signed Claude, bottom right. The same identity slip every Qwen build made, one file out of four this time. The 1-bit M ran into the hour.

Reasoning high: one test out of twenty

Same five files, the dial at high, 60-minute cap. One test finished: the 1-bit S Ledger at 41.9 minutes and 18,286 tokens, with the same top slip as at low, so a blank page. The other nineteen hit the hour with 47,000 to 96,000 characters of thinking kept, about 23,000 tokens per hour. Six of them had started writing an answer before the cap; none of those answers were runnable. Eruption and Blind Artist produced thinking only, on every file. Memory tracked the low ladder within 2 GB, and the 147 GB file failed to load exactly as before.

Reasoning max: one word, on repeat, for three hours

Max is the model's default, so I owed it a try on the 93 GB file. With the usual cap and 65k context: zero of four finished, thinking only, 82,000 to 93,000 characters each. So I gave it a 3-hour cap and a 262k context, which loaded fine at 105 GB peak, no swap, no OOM: zero of four again, 224,000 to 236,000 characters of thinking, zero answer characters.

The transcripts show what happened. After 70,000 to 85,000 characters, about 20,000 tokens, the thinking collapses into one repeated Russian word, on every test, and 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 first cap hit. More time and more context changed nothing. That is not the model being dumb; GLM-5.3-Flash through the API is excellent. That is what 1-bit compression plus this hardware plus a one-hour budget plus no harness looks like: the worst possible case for a frontier open-weight model, and I built it on purpose to find the edge.

Verdict

Great model, wrong hardware. On a 128 GB Strix Halo, the 93 GB 1-bit S at reasoning low is the only combination that finishes the exam, and low is the level this model is least known for. High finishes the Ledger only. Max finishes nothing. The bigger files fit, and the 2-bit XL is the one to pick if you care about the Ledger-style work and can wait, but none of them change the speed, and 9 tokens a second is slower than you read.

Unsloth says the branch gained MTP and faster decoding on September 4, after these runs, up to 3.3x at long context on paper. If that lands anywhere near 20 to 30 tokens a second and high stops breaking, the same ladder on the new branch is the follow-up. Until then I am stopping here with GLM-5.3-Flash on this box. Comfortable use of this model needs something like a couple of DGX Sparks or a big M3 Ultra, and this is all I own.

If you run GLM-5.3-Flash on anything, post the quant, the memory and the tokens per second. They go on the page.

My setup

Gear in this one

Some links on this page are affiliate links. If you buy through them I may earn a small commission at no extra cost to you. It helps support the work. As an Amazon Associate I earn from qualifying purchases.