⚠️ There is a better build of this model — ThakiCloud/Qwen3.8-27B-NVFP4-FP8ATTN
This checkpoint quantizes only the MLP and leaves every attention projection at
bf16. Our mixed build puts attention at FP8 instead, and on the same B200 harness it is
smaller and faster at once: 22.90 GB against 30.14 GB, and 1.641x over bf16 at
saturation against this build's 1.488x (1.675x with --kv-cache-dtype fp8). Neither
build separates from bf16 on paired MMMU.
This repo stays up because the pair is the evidence for a calibration null result, and
because the six-arm comparison needs it. For serving, take the mixed build.
NVFP4 (4-bit weights and activations) quantization of the MLP stack of Qwen/Qwen3.8-27B, produced with
llm-compressor using a GPTQ oneshot over 1,024 calibration samples calibrated on
text only. 30.14 GB on disk against the 55.59 GB bf16 original.
What is actually 4-bit here (corrected 2026-08-19)
⛔ An earlier revision of this card said the vision tower was the only thing left at
original precision. That was wrong, and it understated what stayed in bf16 by a factor of
four. The ignore list holds 511 modules, and only 110 of them are the vision tower:
kept at bf16
count
model.language_model.*.linear_attn.* (48 layers x 7 projections)
336
model.language_model.*.self_attn.{q,k,v,o}_proj (16 layers x 4)
64
lm_head
1
model.visual.* (vision tower)
110
So this checkpoint is NVFP4 on the MLP stack, bf16 everywhere else — every attention
projection in the language model is untouched. That, not the vision tower, is the main
reason the reduction is 1.84x instead of the ~3x a blanket 4-bit pass would give.
The measured numbers below are unaffected: they were taken on this checkpoint as it is.
What changes is what you should call it. If you are comparing against
RadixArk/Qwen3.8-27B-NVFP4, note that theirs is a different recipe — modelopt
mixed precision that puts all 208 attention projections at FP8 and the MLP at NVFP4
with an FP8 KV cache, rather than leaving attention at bf16. Those two checkpoints are not
the same thing under the same name, and a size or score comparison between them is a
comparison of recipes.
This repo is one half of a pair
ThakiCloud/Qwen3.8-27B-NVFP4-GPTQ-mm is the same model, same recipe, same bit width,
same file size — calibrated on text and images instead. Nothing else differs. quantize_meta.json
in each repo records the difference as multimodal: false.
The pair exists because of a result we got on a different vision-language model: with the
recipe held fixed, moving from a text-only calibration set to one that includes images
moved MMMU by 1.89pp on that model. On this one the two calibrations are
indistinguishable (see Measured) — so the pair is published as a null result rather
than a winner. Publishing only one would have implied a difference we did not find.
Measured
Single B200. Serving config identical across every arm and echoed by the harness rather
than assumed: max_num_seqs=256, max_model_len=32768, gpu_memory_utilization=0.90,
CompilationMode.VLLM_COMPILE + CUDAGraphMode.FULL_AND_PIECEWISE.
Throughput — 1.44x at batch 1, still 1.49x at saturation
Ladder 1/8/32/128, three repeats per level, median. Output tokens/s, in=2048 out=256.
concurrency
Qwen/Qwen3.8-27B (bf16)
this build
speedup
1
86.5
126.3
1.46x
8
565.4
814.4
1.44x
32
1,382.4
2,013.1
1.46x
128
2,141.4
3,186.2
1.49x
The gain holds at saturation, which is where serving actually lives — a single-stream
figure alone would not tell you that. The sibling build, calibrated separately, lands
within 0.002 of these ratios at every level.
Quality — no measurable change against bf16
MMMU validation, multiple-choice only (53 open-ended items skipped rather than guessed at),
max_new_tokens=16384, temperature 0, one prompt shared by every arm. Scored by code, not
by an LLM judge. Paired: the table below is restricted to the 246 items that produced a
verdict in all three arms, because arms lose different items to truncation and unpaired
accuracies are not differenceable.
build
MMMU-val MC (246 paired)
vs bf16
McNemar
Qwen/Qwen3.8-27B (bf16, 55.59 GB)
0.8455
—
—
this build (30.14 GB)
0.8577
+1.2pp
p = 0.453 (7 discordant pairs)
There is no detectable quality difference. Nor is there one between the two
calibrations. Read that as "quantization did not break it", not as a gain — 4-bit does not
add knowledge, and a difference this size is well inside the noise at n=246.
⚠️ n=246 can show the absence of a large effect, not of a small one. MMMU only.
⛔ An earlier revision of this card listed bf16 MMMU at 0.4511. That number is withdrawn.
It came from a harness that capped generation at 512 tokens on a model that spends 12,000+
characters reasoning before it answers, so it scored truncated reasoning as wrong answers.
Chance on 4-way multiple choice is 0.25; a "score" near it should have been read as a broken
measurement, and for a day it was not. The same run had a judge-scored task return exactly
0.0 — lmms_eval/tasks/mmmu/utils.py converts a judge failure into judge_score = 0 with
nothing but a log line, so a missing judge server is indistinguishable from a model that
gets everything wrong. Both numbers are gone; these replace them.
Answer extraction failed on 0 items across all three arms.
Quality — math reasoning does move (added 2026-08-22)
The MMMU result above stands, and so does its caveat: n=246 cannot see a small effect.
We went looking on other axes and found one.
GSM8K, full 1,319-item set, greedy, vLLM v0.27.1 server + lm-eval 0.4.12,
identical serving config on both arms:
build
strict-match
flexible-extract
Qwen/Qwen3.8-27B (bf16)
0.5413
0.5451
this build
0.4936 (−4.77pp)
0.5239 (−2.12pp)
Two-proportion test on strict-match: z = −2.45, p = 0.014. An independent rerun of the
same comparison returned −4.32pp (p = 0.026), so this is not a single-run artifact.
Flexible-extract moves −2.12pp (p = 0.275), inside noise. That the gap appears under strict
but not flexible extraction points at format adherence rather than answer correctness —
but that reading comes from the difference between two metrics and is not itself measured.
Korean and general reasoning show no such loss (n = 600 per task, same harness):
task
bf16
this build
delta
p
HAERAE
0.7709
0.7608
−1.01pp
0.680
KoBEST
0.6770
0.6745
−0.25pp
0.926
belebele-ko
0.9000
0.9083
+0.83pp
0.625
ARC-Challenge
0.5700
0.5750
+0.50pp
—
IFEval prompt-strict
0.3050
0.3142
+0.92pp
—
So the honest summary is narrower than "no measurable change". On MMMU, Korean,
commonsense reasoning and instruction following we cannot distinguish this build from bf16.
On GSM8K we can, and it is worse. If your workload is math or multi-step arithmetic
reasoning, measure before you switch.
⚠️ kmmlu_direct and humaneval were run in the same batch and are excluded: kmmlu
scored 0.00–0.12 across every subject and humaneval returned exactly 0.0000 on all three
arms, which we read as parsing/format failure rather than accuracy. gpqa was not run — it
is gated on the Hub.
Speculative decoding on this checkpoint (added 2026-08-23)
We served this checkpoint with the DFlash2 drafter (z-lab/Qwen3.8-27B-DFlash2, K=7)
on a vLLM nightly build and measured two things.
Quality is unchanged. Holding image, checkpoint and task set fixed and toggling only the
drafter, none of 34 metrics separated — GSM8K, IFEval, HAERAE, KoBEST and belebele-ko
included, minimum p = 0.774 (GSM8K strict −0.83pp at p=0.774; six metrics identical to the
decimal). That is what rejection sampling should do, but it is worth measuring rather than
assuming: the losslessness is a property of the algorithm, not automatically of a build.
⛔ The engine dies on very long prompts, and the failure is not graceful. A length ladder
on a max_model_len=1000000 endpoint handled 8,410 / 66,968 / 148,503 / 244,689 tokens
cleanly (acceptance length steady at 1.42–1.47), then returned HTTP 500 at roughly 300k and
killed the EngineCore process — CUDA error: cudaErrorIllegalAddress
(CUDAGraphsC10Utils.h:73). One request takes the endpoint down.
If you enable a drafter here, cap max_model_len at or below ~245k. We set 245,760 and
re-verified: 240,503 tokens serve normally and anything larger returns a clean HTTP 400
instead of crashing. We have not bisected the exact boundary between 244,689 and ~300k.
The drafter also costs 25% of the KV pool — 1,808,112 → 1,354,786 tokens under otherwise
identical settings, so concurrency for million-token requests falls from 1.81 to 1.35. In
long-context serving the drafter's price is capacity, not latency.
Update (2026-08-31). A follow-up run pinned down two more things about this drafter.
Launch config, for reproducibility: z-lab/Qwen3.8-27B-DFlash2 as the draft model, served on
vLLM 0.28.0 with
--speculative-config '{"method":"dflash","model":"z-lab/Qwen3.8-27B-DFlash2","num_speculative_tokens":7}' --mamba-cache-mode align.
The KV overhead is a proportion of the pool, not a fixed token count. Switching
--kv-cache-dtype fp8 on the target checkpoint does not shrink the drafter's relative cost —
it stays in the 34–36% range regardless of KV dtype (measured separately from the 25%
figure above; both point the same direction, that the overhead scales with pool size rather
than being absorbed by a cheaper KV representation).
The speedup does not automatically transfer to other target checkpoints. On the vendor's
own FP8 build, the same drafter helps across the whole concurrency range we tested (c=1–8). On
a third-party INT4 build (RedHatAI), it loses throughput at c=8 — 0.82x of baseline, an 18%
hit. We have not re-run that same ladder on this checkpoint, so treat the 34-metric
quality-neutral result above as specific to this build, and re-verify before assuming a
drafter's win on one target checkpoint carries over to another.
Rebuild noise floor — why the GSM8K gap above is not attributable (added 2026-09-01)
We requantized this exact model with this exact recipe and this exact calibration set a
second time, changing nothing except letting GPTQ run again. That second run is not a
different model — it is the same instructions executed twice — so whatever daylight opens up
between the two builds is the floor under any GSM8K comparison made with this recipe, not a
property of the model.
metric
rebuild-to-rebuild gap (floor)
GSM8K (en) strict-match
3.56pp
GSM8K (en) flexible-extract
2.27pp
gsm8k_ko strict-match
2.13pp
gsm8k_ko flexible-extract
1.37pp
RAG absent-context fabrication rate
27.1pp (7.5σ)
Rule we use: an observed gap has to clear roughly 2x its metric's floor before we call it
attributable to the treatment (here, leaving attention at bf16). Below that, a low p-value
does not rescue it — the p-value assumes the only source of variance is sampling noise on a
fixed model, and the floor shows there is a second, unmodeled source: which quantization run
you happened to get.
Applying that to the GSM8K numbers on this checkpoint:
measurement
gap
floor
ratio
verdict
full set (n=1,319) strict-match, above
−4.77pp
3.56pp
1.34x
not attributable
n=600 subsample (quoted in the -FP8ATTN card's comparison table)
−5.50pp (p=0.056)
3.56pp
1.55x
not attributable
flexible-extract, above
−2.12pp
2.27pp
0.93x
already flagged as noise
Neither GSM8K number for this checkpoint clears the bar. That does not mean "leaving attention
at bf16 doesn't hurt GSM8K" — it means we cannot tell from this data whether it does. The
sibling ThakiCloud/Qwen3.8-27B-NVFP4-FP8ATTN, which quantizes attention as well, shows a
larger gap (−8.33pp) that does clear the same floor, by 2.34x — see that card for the
reading that number supports.
Kernel path
A 4-bit checkpoint can hold NVFP4 weights and still expand them to bf16 for every matmul —
slower than bf16, saving only memory. Verified on B200 (SM100) with vLLM 0.27.1 that this
build reaches a native FP4 kernel:
⚠️ Blackwell only. On Hopper (H200/H100) vLLM falls back to weight-only Marlin
emulation, which we measured at roughly 0.85x of bf16 on a comparable model. NVFP4 is not
a speed win on SM90.
Does this fit your GPU? (added 2026-09-01)
A separate same-run kernel-throughput comparison on this model family (single B200) put a
plain NVFP4 arm at 3.52x the ceiling throughput of a W4A16 (Marlin) arm — 3,597.9 vs 1,022.8
output tok/s — with 3.12x better tok/J and 5.1x lower time-to-first-token (2.093s vs 10.664s).
That is the case for NVFP4 on hardware that can hold it.
The catch is exactly what this card is about: this build's 30.14 GB does not leave much
headroom on a 32 GB-class card once you account for KV cache and runtime overhead. On that
class of card, W4A16 is the build that actually fits; on a 96 GB-class card, an NVFP4-family
build is worth the squeeze for the throughput and tok/J win above — and at that point you
should reach for ThakiCloud/Qwen3.8-27B-NVFP4-FP8ATTN (22.90 GB) rather than this one, per
the banner at the top of this card.
quantize_meta.json ships in this repo: GPTQ oneshot, 1,024 samples, 2367 s, and the
calibration flag that distinguishes this build from its sibling. Check the card against it.
Scope
One model, one recipe, one engine version, one GPU generation for the kernel claim.
Evaluated on MMMU, ChartQA, GSM8K, IFEval, HAERAE, KoBEST, belebele-ko and ARC-Challenge.
Code and video behaviour are unevaluated and unclaimed; kmmlu_direct and humaneval were
run and excluded as format failures, and gpqa is gated on the Hub and was not run.
Long-context is covered on the speed axis only (the drafter ladder above), not on quality.
