Mia's 2-Spark recipe, version 1104, commit d4226a1: GLM-5.3-Flash

How fast it runs

Long prompts

Method: code+prose; 2 run(s) each; 512 tokens; 1 warm-up(s); median; context steps: prompts of exactly 16K, 32K, 64K and 128K tokens (a filler log first, the code or prose task prompt last), 2 runs per step, 512 output tokens; decode = tokens after the first / time after the first token; prefill = prompt tokens / time to first token of the same request, median over its requests; requests at once: C1,C2,C4,C8 at once; 2 rounds each; 512 tokens per request; prose prompt; total = all tokens / wall time, per request = tokens / end-to-end time; median over the rounds. Measured 2026-11-10. Raw files: speed result · speed contract · log · long answer.

Requests at once: the server has 8 slots (from the author's recipe). The speed test ran with Mia's recipe, version 1104 (MiaAI-Lab/GLM-5.3-Flash-DGX-Sparks-1104), in its state of commit d4226a1 of Nov 9, 2026 (112 patches). The earlier page of this recipe version, commit 3f44ec2 with 105 patches, measured on Nov 4 and 5, 2026, stays online; it is another run and none of its numbers are mixed into this page. The speed test ran on TensorFold's native engine (commit f350cd3e, Zig 0.17.0) with the recipe's 112 patches, tensor parallel over two boxes (CX7 RDMA, 109.2 and 109.3 Gb/s on the two links before the run), DFlash2 drafts and vision on as the recipe sets them, a context window of 1,048,576 tokens and 8 requests at once (PARALLEL=8 in the local.sh of her served pair, README line 173; the script default is 4). I used her served pair's settings, with top-k 20 as her pair sets it (README line 213), and lowered nothing. Her worker-weights check wants a marker file with the revision of her pin (76c0b517); my copy of the weights is snapshot 078455ff, whose 83 weight files have the same sha256 as her pin (compared with the Hugging Face file lists; only the readme and licence files differ), so the marker with her pin revision is on both boxes. Sampling is the model card's plus top-k 20, the same with thinking on and off: temperature 1.0, top_p 0.95; seed 42; the request flag for no thinking; no cached tokens in the 61 requests that were checked. The headline 39.3 tok/s is the mean of the code prompt (37.2) and the prose prompt (41.3), medians of two runs each (spread 0.0 % and 3.8 %). Drafts were accepted at a rate of 55.7 % (1,092 of 1,962 drafted tokens), 2.1 tokens per step on average. The server has 8 slots, so the four concurrency levels (1, 2, 4 and 8 at once) are all inside its slot count. The long answer ended after 8,367 tokens with the reason "stop", not at a limit, after 149.3 s. The prefill of the first context step (16K tokens) is 1422.6 tok/s, that of the three later steps 1599.7 to 1703.2 tok/s; I did not look into why. Mia's README lists decode numbers for prose at temperature 0.7 with thinking on, measured with sparkDash on this pair. Speed v2 uses other prompts, temperature 1.0 and thinking off, so the numbers are not comparable. The recorded command below says that Mia's own decode numbers were reproduced separately with her method; that separate run is not shown on this page.

Differences from the author's setup, stated openly. (1) Image: I built the image myself from the recipe's tree at commit d4226a1 (112 patches, tag ef2e6847aa4a). It is not the image named in her README (8c4ca7bd8c39, 105 patches, line 63), and her changelog of Nov 9 says that a rebuilt image is not yet measured on her pair. (2) Disk check: one of the two boxes had 16 GiB of free disk, and the recipe's start script asks for 40 GiB for the image (IMAGE_FREE_GIB, scripts/config.sh line 252). I started with IMAGE_FREE_GIB=10 (its start log says about 10 GiB are needed). The variable only changes this free-disk check before the build, but it is not part of the start text below. (3) Prompt cache: at the start of this session the server reported a prompt cache of 1.8 GiB on each of the two ranks and 4,098 pages, with 17.1 GiB of room left after the 9.5 GiB reserve. In the LiveCodeBench session it reported 2.0 GiB (room 17.3 GiB); the start line of the memory-test session was not saved. Mia's README names 3.6 GiB for her pair (line 143), her changelog records startup lines with 3.1 GiB (Nov 7), and it says that the cache is what the room leaves after the streams' windows and that the room moved between 17.8 and 18.9 GiB across starts with no config change (line 54). I did not find why my cache is smaller. My suspicion, which I did not show, is that more memory was in use on my boxes at the start than on hers. I changed no system setting. The speed test sends every prompt cold (none of the 61 checked requests had cached tokens), so it does not use the cache; a multi-turn chat that reuses earlier turns is not measured here. (4) My first start of the recipe on these boxes, before the sessions on this page, stopped: the kernel log of the first box shows "NVRM: Out of memory" at the end of the model load, and my memory guard stopped the server. I did not find the cause and changed no system setting; the sessions on this page started afterwards with the same settings.

The load guard (a sample every half second) repeated 2 of its 38 steps: the first round of the 8-at-once step, because the kernel's memory-compaction thread (kcompactd0) used up to 82.0 % of one core for 3 samples in a row, and the second code run at 64K context, because the snapd service used up to 52.0 % for 2 samples in a row. Both were run again after 5 seconds of quiet, and the result file marks the 64K code step as noisy. In three kept steps a peak above 25 % was recorded (kcompactd0 at 75.8 %, 46.1 % and 64.0 %) that did not last two samples in a row, so the guard did not repeat them. The quiet check (one sample every 10 seconds, 143 samples) saw at most 1.8 % of one core, the test driver itself.

python3 speed2.py --base-url http://127.0.0.1:35031/v1 --label glm-mia-d4226a1-1010-speed-r0 --prompts $BENCH/repos/rigmark-v2/prompts-v2.json --output $RUN/glm-mia-d4226a1-1010-speed-r0/glm-mia-d4226a1-1010-speed-r0/speed2 --contract $RUN/glm-mia-d4226a1-1010-speed-r0/glm-mia-d4226a1-1010-speed-r0/speed2/speed-contract.json --parts decode,prefill,conc,l1,ctx --prefill-depths 2048,8192,32768 --prompt-ids code,prose --runs 2 --max-tokens 512 --warmups 1 --prefill-runs 3 --conc-levels 1,2,4,8 --conc-rounds 2 --conc-max-tokens 512 --conc-prompt prose --ctx-depths 16384,32768,65536,131072 --ctx-runs 2 --ctx-max-tokens 512 --tokenizer-json $HOME/models/glm53-mia-4bpw-1104/tokenizer.json --tokenizer-python $HOME/not published --sampling {"temperature": 1.0, "top_p": 0.95, "top_k": 20} --sampling-source Checkpoint generation_config.json:9-10 (temperature 1.0, top_p 0.95; Mia-AiLab/GLM-5.3-Flash-EXL3-4bpw-TensorFold @078455ff = pin 76c0b517 weights, copy $HOME/models/glm53-mia-4bpw-1104/generation_config.json) plus the pair's server default top_k 20 (SERVE_EXTRA_ARGS in scripts/local.sh, README.md:213, CHANGELOG.md:27; applies to requests that send no top_k). Mia's own official decode numbers use T=0.7 with thinking ON (README.md:66) and are reproduced separately (mia_method.py, own column, never mixed). Speed v2 runs with thinking off (enable_thinking false per request), think part with thinking on. --model GLM-5.3-Flash --no-cache-prompt --server-log $RUN/server-A.log --engine-version-file $RUN/engine-version.txt --server-slots 8 --server-slots-source MiaAI-Lab/GLM-5.3-Flash-DGX-Sparks-1104@d4226a1 scripts/local.sh PARALLEL=8 (README.md:173; performance table README.md:80-86 '8 streams'; scripts/config.sh:139 default 4, 'auto: up to 8'); rank 0 start log shows '8 at once' --server-ctx 1048576 --server-ctx-source card GB10 box 3-glm53-flash-tf-zig-1104-2x-mia-d4226a1 field ctx

How well it answers

Thinking off, except the last test. Seed 42. Sampling from the model card. A cut-off answer counts as wrong. Measured 2026-11-10.

Note from the run: MMLU-Pro and GPQA Diamond ran with an answer limit of 1,024 tokens (my usual limit for these two tests is 48; the card of this setup sets 1,024, which is the limit Mia's sparkDash uses for MMLU); same questions, seed 42, first-letter extraction. 1 of 400 MMLU-Pro answers and 7 of 198 GPQA Diamond answers were still cut off at that limit and count as wrong. 0 of 200 IFEval answers were cut off at their limit of 4,096 tokens, 12 of 80 long-text answers at their limit of 128 tokens and 3 of 30 GPQA answers with thinking at the limit of 16,384 tokens; these count as wrong too. 13 of 40 LiveCodeBench answers were cut off at their limit of 32,768 tokens and count as wrong (job glm-mia-d4226a1-1010-lcb, run after the other tests); the median LiveCodeBench answer is 6,529 tokens long. The questions of each test ran 8 at a time, the slot count of the recipe's server (same questions, limits and seed for every question). Sampling: temperature 1.0, top_p 0.95 (the model card) and top-k 20 (the served pair's server default). The GPQA test with thinking ran at the recipe's default reasoning effort (xhigh, scripts/config.sh line 213); no request sets it. Speed v2, MMLU-Pro, GPQA Diamond, IFEval, the long-text tests and GPQA with thinking ran one after the other on Nov 10, 2026 on one freshly started server; LiveCodeBench and the memory test each ran in their own session on a freshly started server. These quality tests have no foreign-load guard; their sampling files list the busiest other process every 10 seconds (the numbers below are read from them): in the four instruction tests 3 of 251 samples were above 25 % CPU (52.0 to 57.1 %, a firmware query), in the thinking test 1 of 173 (100 %, a container status query), and in LiveCodeBench 8 of 629 (up to 100 %: 3 package-manager checks, 4 firmware queries and 1 sample from the test's own code runner). The answers files of MMLU-Pro, IFEval and LiveCodeBench are not published: the site's filter flagged text in the model's answers (a physics calculation with power units in MMLU-Pro, a word of a course outline in IFEval that looks like a host name to the filter and the letter W used as a label in a puzzle in LiveCodeBench; none of them is a power reading or a host name). The rows were counted from the complete files, and the summary, the log and the contract of each of these tests are published. The 10-second sampling files of these tests and the sample file of the speed test's load guard are not published either (the filter blocks them, among other things for the name of a network-service process). These are my own tests, and no other run of this setup is mixed into the numbers on this page.

bash tools/quality2/measure_all.sh tools/quality2/cards/glm53-flash-tf-zig-1104-2x-mia-d4226a1.json --parts mmlupro --job glm-mia-d4226a1-1010

bash tools/quality2/measure_all.sh tools/quality2/cards/glm53-flash-tf-zig-1104-2x-mia-d4226a1.json --parts gpqa --job glm-mia-d4226a1-1010

bash tools/quality2/measure_all.sh tools/quality2/cards/glm53-flash-tf-zig-1104-2x-mia-d4226a1.json --parts ifeval --job glm-mia-d4226a1-1010

bash tools/quality2/measure_all.sh tools/quality2/cards/glm53-flash-tf-zig-1104-2x-mia-d4226a1.json --parts lcb --job glm-mia-d4226a1-1010-lcb

bash tools/quality2/measure_all.sh tools/quality2/cards/glm53-flash-tf-zig-1104-2x-mia-d4226a1.json --parts ruler --job glm-mia-d4226a1-1010

bash tools/quality2/measure_all.sh tools/quality2/cards/glm53-flash-tf-zig-1104-2x-mia-d4226a1.json --parts ruler --job glm-mia-d4226a1-1010

bash tools/quality2/measure_all.sh tools/quality2/cards/glm53-flash-tf-zig-1104-2x-mia-d4226a1.json --parts think --job glm-mia-d4226a1-1010

How much context fits

Memory test. The recipe fixes the context (1,048,576 tokens), so the server is started once with the recipe's own command and reserves its memory up front. Then, for each size (4K, 8K, 16K ... up to the largest size tested), one prompt that fills about 95 % of that size is sent, and the model must answer. The largest size that answered counts; it is the highest size tested, not a limit of the setup. Decode speed at that size was not measured. Peak memory is, on each of the two boxes, the rise of used memory (MemTotal - MemAvailable) over the idle value before the step; the table and the number show the highest of the two boxes. Unified memory, so GPU and CPU share it. Measured 2026-11-10. Raw files: result of every step · log · contract.

Note from the run: fit memory per node, highest rise over idle: box-3 102.1 GiB, box-4 101.5 GiB. The recipe fixes the context (1,048,576 tokens, 8 requests at once), so the server is started once with the author's command and reserves its memory up front. Every size from 4K to 128K was sent as a prompt of about 95 % of the size and answered. The largest size tested is 131,072 tokens, the highest size in my list, not a limit of the setup: the window is 1,048,576. Decode speed at these sizes is not measured here (the memory test returns no timings; the speed section has decode and prefill at 16K to 128K). The memory number is the rise of used memory over the idle value (unified memory, so GPU and CPU share it). The number shown (102.06 GiB) is the highest rise on the first box (at the largest size, 124,518 prompt tokens); the second box peaked at 101.46 GiB. Most of it is the model and the reserved cache, which the server takes at its start. The start line of this session's server was not saved, so its prompt-cache size is not recorded (the two other sessions are on this page: see the note at the speed test).

bash tools/quality2/measure_all.sh tools/quality2/cards/glm53-flash-tf-zig-1104-2x-mia-d4226a1.json --parts fit --job glm-mia-d4226a1-1010-fit

How it plays the Gauntlet

One solo race, season seed 21, one attempt, no help.

Game rank as of Nov 10, 2026. The rank is the order of the Gauntlet finish times in this memory class. Race page and raw files.

The race ran the recipe named on its race page, with the sampling and answer limit of its lane config. The engine build of the race is not recorded in the race files.

Result (result.json) Race log (log.jsonl) Model answers (io.jsonl) System prompt Exact command Verify output Model config Race page

Also made this possible

• ModelZ.ai (GLM-5.3-Flash)Hugging Face @Zai_org GitHub Website
• Draft model (DFlash2, revision bf582e4e)Inco AI@inco_ai Hugging Face GitHub Website
• Engine (TensorFold, commit f350cd3e)Ash Hart@ashxhart GitHub Hugging Face Website
• EXL3 formatturboderp@turboderp_ Hugging Face GitHub

Run it yourself

What you need and how to start it, step by step. The results below come from exactly this setup.

What you need

Hardware

2x NVIDIA GB10 boxes linked by CX7 RDMA (109.2 and 109.3 Gb/s on the two links, measured before the run): 1x Gigabyte AI TOP ATOM and 1x Lenovo ThinkStation PGX. Both run driver 580.178.04 and kernel 7.0.0-1019-nvidia, GPU clock capped at 2200 MHz.

Model

Mia GLM-5.3-Flash EXL3 4bpw (rev 078455ff = weights of pin 76c0b517) + incoai DFlash2 (rev bf582e4e)

Draft model (DFlash2, revision bf582e4e)

huggingface.co/incoai/GLM-5.3-Flash-DFlash2

Engine

Recipe commit d4226a1 of Nov 9, 2026 (112 patches), TensorFold's native engine (commit f350cd3e), tensor parallel on 2 GB10 boxes, EXL3 4 bpw with DFlash2 drafts — the program that loads the model and answers requests · github.com/ashhart/TensorFold

Recipe (version 1104, commit d4226a1) and EXL3 quant

github.com/MiaAI-Lab/GLM-5.3-Flash-DGX-Sparks-1104 by Mia (MiaAI-Lab)

Steps

1. Get the recipe, the engine and the model files from the list above.

2. Start the server with this command. It loads the model into memory. When the model has loaded, the server waits for requests on port 8888.

   Copy

   ./start.sh in glm53-mia-d4226a1 on box-3 via spark-2x-glm-d4226a1.sh start (scripts/local.sh: \
     WORKER=me@127.0.0.1, \
     PORT=8888, \
     PARALLEL=8, \
     CONTEXT=auto, \
     KV=fp8, \
     DENSE=q8, \
     HEAD=q8, \
     ABSORB_MMA=1, \
     HOST_CACHE_GIB=0, \
     MEMORY_RESERVE_GIB=9.5, \
     MODEL_SOURCE=local, \
     WORKER_WEIGHTS=check, \
     SERVE_EXTRA_ARGS='--top-k 20'); head box-3 rank 0, worker box-4 rank 1

   The file names are the ones we used. Run the command in the folder that holds them, or put the full path in front of each name.

TensorFold native Zig engine f350cd3e (ashhart/TensorFold, Zig 0.17.0, NCCL 2.30.x) + 112 patches of Mia's recipe 1104 (d4226a1), TP=2, DFlash2 drafts (4 bit), image tensorfold-glm53-spark:ef2e6847aa4a; ./start.sh in $HOME/glm53-mia-d4226a1 on box-3 via $HOME/glm53-d4226a1-0911/spark-2x-glm-d4226a1.sh start (scripts/local.sh: WORKER=me@127.0.0.1, PORT=8888, PARALLEL=8, CONTEXT=auto, KV=fp8, DENSE=q8, HEAD=q8, ABSORB_MMA=1, HOST_CACHE_GIB=0, MEMORY_RESERVE_GIB=9.5, MODEL_SOURCE=local, WORKER_WEIGHTS=check, SERVE_EXTRA_ARGS='--top-k 20'); head box-3 rank 0, worker box-4 rank 1