Maya GLM-5.3-Flash on an RTX 4090: Qualbench, Speed, and a 32K Ceiling

This host already runs Strata IQ3_S (Qwen3.8-Flash-Next) on the RTX 4090 for Bifrost llm-88. The open question was whether GLM-5.3-Flash, a much larger MoE, can share that card via Project Maya and how it scores on the same text suites.

Not a side-by-side live deploy. Maya needs the whole 4090 and pins most free host RAM. Strata was stopped for the window, then restored. ASR / embeddings on the 5060 Ti stayed up.

Headline:

Stack	Qualbench-50	Wall	Decode (short)	Context
Strata IQ3_S (prod, recheck same evening)	50/50	644 s	Monitor warm 119 t/s; probe ~100-180	262K adv. / ~248K practical
Maya Maya-S24 (this run)	41/50	596 s	~27-45 t/s	32K hard

Seven of Maya’s nine fails are 64K / 150K long-context tasks that the server rejects with HTTP 400. Inside 32K, coding and security look strong.

The stack

	Value
Date	2026-10-30
Engine	Project Maya v1.1.41
Quant	Maya-S24 (~94.7 GB, pre-downloaded)
Served id	`glm-5.3-flash`
GPUs	GPU0 RTX 4090 only (`--no-vision`)
Port	`:8090` (Strata keeps `:8080` for rollback)
Advertised context	32768
Host RAM while serving	pins most free RAM (~85+ GiB used class)
VRAM GPU0	~23.7 GiB

LAN INPUT for 8090 was not opened on this host (only 8080 has a nopasswd helper). Client suites reached Maya through an SSH local forward (localhost → 127.0.0.1:8090).

Smoke and official Maya bench

Gate	Result
`/v1/models`	OK: `glm-5.3-flash`, `n_ctx=32768`
Chat `PONG`	content OK; decode 27.5 t/s
`./maya.sh --bench`	decode mean 30.5 t/s; prefill 2K 389, 8K 1295

Vs Strata Monitor-class decode on the same card (~110-180 t/s on short turns), Maya is roughly 4-6× slower at generation. Prefill can still be competitive on mid-size prompts once warmed.

Monitor (suite)

Cell	prompt_tok	prefill t/s	decode t/s
warm8k	10424	1648	42.4
reuse2k	2624	630	27.3

Agentbench

Same FreeToken-oriented suite used against Strata. Maya is not FreeToken:

Mode	Pass / fail / skip	Wall
fast	49 / 6 / 11	~202 s
slow	49 / 11 / 6	~213 s

Expected fail class:

models payload uses meta.n_ctx / context_length, not max_model_len;
/v1/messages/count_tokens → 404;
some malformed-request cases return 200 where the suite wants ≥400;
FreeToken-specific think_off / tool_choice=required assertions;
slow suite needles at 60K / 200K → context_length_exceeded.

Chat completions, streaming, and tool calling were largely green, matching qualbench mcp-tools 8/8.

Context matrix (≤32K)

Strata’s raw tier C targets ~32K filler but measures ~42K prompt tokens on this harness, over Maya’s hard cap.

Tier	Result	Notes
A ~1K	PASS	cold decode ~36-50 t/s
B ~4K	PASS	cold prefill ~1007 t/s
C target 32768	FAIL HTTP 400	~42.5K prompt
C-calib ~24K target → ~31.1K pt	PASS cold	prefill 2329 t/s, decode 44.8
C-calib warm (session reuse)	PASS	TTFT ~0.06 s

So Maya does hold ~31K prompts on this card; the suite’s naive 32K filler sizing still overflows.

Qualbench-50

Same categories as the Strata article (no multimodal):

Category	Maya	Strata IQ3_S (recheck 2026-10-30)
Total	41/50 @ 596 s	50/50 @ 644 s
java-spring	10/10	10/10
ts-angular	7/8	8/8
sql-migrations	5/6	6/6
mcp-tools	8/8	8/8
security-review	8/8	8/8
long-context	3/10	10/10

Quality fails inside budget: ts-angular/03-shopping-cart, sql-migrations/02-unique-constraint-dupes. Long-context: all 8K needles passed; every 64K / 150K task failed at request time (400), not at retrieval.

Wall time looks “faster” than Strata’s ~11 min only because the long-context category mostly aborted in <1 s each. That is not a throughput win.

What this means

Scenario	Prefer
Prod Bifrost `llm-88` / OpenCode on this host	Strata IQ3_S when restored (engine v0.2.11, `parallel: 2`; gpu-unlock off). Temporary 2026-11-07 window: Gemma-4 QAT on `:8080`
Long prompts / ~150K-248K work	Strata (Maya’s 32K ceiling ends the race)
Try GLM-5.3-Flash locally on one 4090	Maya-S24 (fits, decode ~30 t/s, strong <32K coding/security)
Try Gemma-4 31B QAT on the same card	vLLM W4A16 (~41 t/s, 3072 cap; see /ai/gemma-4-31b-qat-rtx-4090-qualbench)
Concurrent chat + ASR on GPU1	Stop Strata for Maya; do not layer-split Maya with the 5060 Ti in v1
Agent harness written for FreeToken field names	Expect contract fails; OpenAI chat + tools still usable

Lessons

Same suite, different ceiling. A 41/50 with seven context rejects is mostly a product limit, not a “GLM cannot code” story.
Decode SLA matters. ~30 t/s is usable for interactive coding; it is not Strata Flash-Next’s feel on this box.
RAM is part of the cost. Maya pins almost all free host RAM for expert tiers; plan downtime and do not leave it overnight without watching thrash.
Rollback first. Keep Strata on :8080, Maya on :8090, never disable units, only stop, so restore is one systemctl --user start.

Resources

Project Maya: https://github.com/mw00/project-maya
Strata: https://github.com/Niko1221/Strata
llm-tests (qualbench / agentbench): https://github.com/berdachuk/llm-tests
Gemma-4 31B QAT on an RTX 4090: Qualbench, Speed, and a 3K Ceiling
Running Large Language Models Locally: Complete Hardware Guide for GLM-4.7 Deployment

Published on 10/30/2026