Qwen3.8-27B GSQ-RCO — Eight-Quant Benchmark

The verdict

The fastest file was not automatically the best broad-use model. Quality was weighted around the actual jobs this local stack needs to perform; HumanEval+ contributed only 10% of the balanced index.

Best general-purpose choice

IQ3_S MTP

It produced the highest balanced score, nearly halved the adaptive long-task time versus IQ3_S, and led the Researcher and Coder role scores. The tradeoff is memory: only 1,673 MiB remained free in the representative run.

83.19balanced index

Fastest adaptive runIQ2_XS MTP · 48.29s

The speed-first choice, with a 32.7% elapsed-time reduction over its base pair.

Best non-MTP choiceIQ3_S · 81.34

The highest base-model balanced score and the top C-Suite role score.

Watch the findings

The 96-second explainer covers the experimental setup, $2.95 cost, MTP speed and memory tradeoffs, quality results, 64K context, and role-specific recommendations.

ISTA-DASLab Qwen3.8-27B Quant BenchmarkOpen on YouTube ↗

Best model by job

One universal winner is less useful than a routing policy. These scores reweight the same evidence for five distinct jobs in a private AI staff.

C-Suite work

IQ3_S75.03

Researcher / Scout

IQ3_S MTP89.51

Hermes Coordinator

IQ2_S MTP91.05

Librarian / Knowledge

IQ2_S MTP90.16

Forge / Coder

IQ3_S MTP88.88

All eight results

Switch the chart between quality, coding, elapsed time, and remaining memory. Lower is better for time; higher is better for the other views.

Model	Balanced	HE+ pass@1	Adaptive time	Peak VRAM	Free VRAM	Selected profile
IQ3_S MTP WINNERQwen3.8-27B GSQ-RCO	83.19	89.63%	53.49s	14,176 MiB	1,673 MiB	MTP 3 · ubatch 256
IQ3_SQwen3.8-27B GSQ-RCO	81.34	87.80%	105.79s	13,286 MiB	2,563 MiB	MTP 0 · ubatch 1024
IQ3_XXSQwen3.8-27B GSQ-RCO	81.25	90.24%	98.51s	11,692 MiB	4,159 MiB	MTP 0 · ubatch 1024
IQ2_XS MTPQwen3.8-27B GSQ-RCO	80.80	90.24%	48.29s	11,270 MiB	4,581 MiB	MTP 3 · ubatch 512
IQ2_S MTPQwen3.8-27B GSQ-RCO	80.66	85.98%	49.62s	12,120 MiB	3,731 MiB	MTP 2 · ubatch 1024
IQ2_SQwen3.8-27B GSQ-RCO	80.04	84.76%	65.45s	10,896 MiB	4,955 MiB	MTP 0 · ubatch 1024
IQ2_XSQwen3.8-27B GSQ-RCO	79.12	85.98%	71.73s	10,222 MiB	5,629 MiB	MTP 0 · ubatch 1024
IQ3_XXS MTPQwen3.8-27B GSQ-RCO	78.68	87.80%	49.76s	13,066 MiB	2,785 MiB	MTP 3 · ubatch 1024

What MTP changed

MTP delivered meaningful acceleration in every matched pair, but it always consumed more VRAM and did not guarantee better broad-use quality.

IQ2_XS

+48.5% faster

Adaptive time: 71.73s → 48.29s
Elapsed reduction: 32.7%
Balanced index: +1.68
Additional VRAM used: 1,048 MiB

IQ2_S

+31.9% faster

Adaptive time: 65.45s → 49.62s
Elapsed reduction: 24.2%
Balanced index: +0.62
Additional VRAM used: 1,224 MiB

IQ3_XXS

+98.0% faster

Adaptive time: 98.51s → 49.76s
Elapsed reduction: 49.5%
Balanced index: −2.57
Additional VRAM used: 1,374 MiB

IQ3_S

+97.8% faster

Adaptive time: 105.79s → 53.49s
Elapsed reduction: 49.4%
Balanced index: +1.85
Additional VRAM used: 890 MiB

How to read “faster”: speedup is calculated as base time divided by MTP time, minus one. The separate elapsed-reduction figure answers the more intuitive question: how much less wall-clock time did the run take?

Quality beyond coding

The applied suite used 24 synthetic tasks per model across six practical categories. This prevented 164 coding problems from dominating a model intended for broad research, operations, and knowledge work.

Applied capability scores — all models

Model	Business	Research	Knowledge	Agent tools	Sysadmin	Executive content
IQ2_XS	47.92	95.83	96.88	80.00	78.75	64.58
IQ2_XS MTP	56.25	95.83	96.88	80.00	78.75	60.42
IQ2_S	56.25	95.83	100.00	71.67	78.75	64.58
IQ2_S MTP	38.54	100.00	100.00	95.83	78.75	60.42
IQ3_XXS	57.29	92.71	96.88	75.83	83.75	72.92
IQ3_XXS MTP	51.04	92.71	96.88	75.83	78.75	64.58
IQ3_S	59.38	100.00	96.88	80.00	73.75	56.25
IQ3_S MTP	56.25	100.00	96.88	90.83	78.75	60.42

HumanEval+ counts — 164 problems per model

Model	Evaluated	HumanEval passed	HumanEval+ passed	HE+ pass@1
IQ2_XS	164	150	141	85.98%
IQ2_XS MTP	164	155	148	90.24%
IQ2_S	164	145	139	84.76%
IQ2_S MTP	164	145	141	85.98%
IQ3_XXS	164	156	148	90.24%
IQ3_XXS MTP	164	150	144	87.80%
IQ3_S	164	151	144	87.80%
IQ3_S MTP	164	152	147	89.63%

Role-weighted scores — all models

Model	C-Suite	Researcher	Coordinator	Librarian	Coder
IQ2_XS	68.89	85.80	85.31	85.95	85.22
IQ2_XS MTP	72.84	86.22	84.90	85.74	87.35
IQ2_S	72.69	86.17	83.53	85.79	83.67
IQ2_S MTP	66.87	89.27	91.05	90.16	88.11
IQ3_XXS	72.95	85.70	85.15	86.05	87.57
IQ3_XXS MTP	69.41	84.24	83.53	84.49	85.35
IQ3_S	75.03	87.78	84.73	86.31	85.34
IQ3_S MTP	74.76	89.51	88.95	88.95	88.88

64K context

Long context was treated as a real deployment requirement, not a launch-parameter checkbox.

8/8completed the long-context workload

The fixed workload contained roughly 56,976 content tokens and ran inside a 73,728-token allocation.

What the test establishes

Every quant completed the same long Hermes-style task at the target class needed for a 64K deployment.
Every file also successfully allocated context through 131,072 tokens on the 16GB test GPU.
Flash Attention and Q4_0 K/V cache were held as the approved long-context baseline.
The largest MTP model still fit, but IQ3_S MTP left only 1,673 MiB of representative free VRAM.

Important: a successful 131K allocation probe is not the same as completing a 131K-token reasoning workload. The strongest end-to-end evidence here is the approximately 57K-token content task; 131K should be described as allocation capacity.

Methodology

Four Vast.ai shards ran matched base/MTP pairs on the same GPU class and software baseline. Scoring was completed in an isolated local Docker sandbox.

Controlled platform

NVIDIA RTX 5060 Ti 16GB

GPU VRAM reported: 16,311 MiB
CPU: AMD EPYC 7742, 64 threads
Driver: 595.91.07
llama.cpp: 0.5.0-dev
Commit: c8296709920f9c1ae168bfd5fe66f9f73637bd60
Four shards; one matched quant pair per shard
Final benchmark rental cost: $2.95

Balanced-index weights

Broad use beats one benchmark

Business analysis20% Research20% Knowledge work20% Agent / tool behavior15% Executive content10% HumanEval+10% Practical sysadmin5%

How the suite was scored

Each model produced three repeats of the fixed long-context workload plus tuned adaptive runs. The applied suite contributed 24 synthetic outputs per model across six categories. HumanEval+ contributed 164 problems per model. Role scores reweighted these same categories for C-Suite work, research, coordination, knowledge management, and coding. Speed and memory were reported separately so a fast but weaker response could not win simply by finishing first.

Limitations and interpretation

This is a controlled comparison of these eight files on one GPU class, not a universal model leaderboard. Applied-task samples are intentionally practical but finite; small score differences should be treated as directional. Rental prices vary over time. MTP gains depend on draft-token acceptance, workload, and tuning. The 131K result is an allocation probe, not a completed 131K active-token workload.

Decision rule

Choose IQ3_S MTP when one model must cover many jobs. Choose IQ3_S for C-Suite analysis without MTP overhead, IQ2_S MTP for coordinator and librarian roles, and IQ2_XS MTP when speed and headroom matter most. In a multi-model private AI staff, route by role rather than forcing one quant to do everything.

Eight quants.One practical decision.

The verdict

IQ3_S MTP

Watch the findings

Best model by job

C-Suite work

Researcher / Scout

Hermes Coordinator

Librarian / Knowledge

Forge / Coder

All eight results

What MTP changed

IQ2_XS

IQ2_S

IQ3_XXS

IQ3_S

Quality beyond coding

64K context

What the test establishes

Methodology

NVIDIA RTX 5060 Ti 16GB

Broad use beats one benchmark

Eight quants.
One practical decision.