GAUNTLET Bench
← Leaderboard

Qwen3.8-27B MLX 4bit (lmstudio-community)

Tested Rank #23 of 53 · 7/8 GAUNTLET progress
Generalist — 82.9 / 100 G Agentic — 100 / 100 A Understanding — 91.3 / 100 U Needle — 91.2 / 100 N Thinking — pending T Live — 78.1 / 100 L Engineering — 99 / 100 E Throughput — 6.9 / 100 T

Hollow markers & dashed spokes: axis not yet scored

G
82.9
A
100
U
91.3
N
91.2
T
n/d
L
78.1
E
99
T
6.9

Specification

Parameters
27B dense
Architecture
qwen3_5
Size on disk
15 GB
Quantization
4bit
Format
MLX
Runtime
LM Studio
Class
Standard
Reasoning (CoT)
Yes — emits reasoning tokens
Internal ID
M25h
Mean speed
23.8 tok/s across suites
Model card
huggingface.co

Suite results

Suite Score Avg / 20 Tests n tok/s Run
General capability (13-task real-workload suite) 1a 215.6 / 260 19.6 11/13 ● n=2 25.5 run 2026-09-07
Test Category Score tok/s
A1 Dev/Ops Scripting 20/20 25.1
A2 Dev/Ops Scripting 20/20 27.8
A4 Dev/Ops Scripting 19/20 25.3
A5 Dev/Ops Scripting 20/20 17.1
A6 Dev/Ops Scripting 20/20 26.7
B1 Document Processing 20/20 24.8
B2 Document Processing 20/20 27.6
C1 Content Production 20/20 19
C2 Content Production 19/20 26.5
C3 Content Production 20/20 29.2
C4 Content Production 19/20 31.2

Category mean Dev/Ops Scripting 19.8Document Processing 20Content Production 19.5

Agentic tool-calling & protocol adherence 1b 160 / 160 20 8/8 n=1 24.1 run 2026-09-02
Test Category Score tok/s
WA1 Tool Calling & Protocol Adherence 20/20 24.9
WA2 Tool Calling & Protocol Adherence 20/20 25.4
WA3 Tool Calling & Protocol Adherence 20/20 25
WA4 Tool Calling & Protocol Adherence 20/20 23.5
WA5 Tool Calling & Protocol Adherence 20/20 20.8
WB1 Agent Robustness & State Management 20/20 18.4
WB2 Agent Robustness & State Management 20/20 19.3
WB3 Agent Robustness & State Management 20/20 19.9

Category mean Tool Calling & Protocol Adherence 20Agent Robustness & State Management 20

Coding depth 1c 178.2 / 180 19.8 9/9 n=1 24 run 2026-09-27
Test Category Score tok/s
K1 Coding — Concurrency 20/20 29.8
K2 Coding — Algorithms 20/20 27
K3 Coding — Security Review 18/20 25.3
K4 Coding — Performance 20/20 23
K5 Coding — Refactoring Under Constraints 20/20 22.2
K6 Coding — Test Writing 20/20 24.3
K7 Coding — React Component (typed) 20/20 22.7
K8 Coding — TypeScript Type System 20/20 19.9
K9 Coding — React Debugging 20/20 21.5

Category mean Coding — Concurrency 20Coding — Algorithms 20Coding — Security Review 18Coding — Performance 20Coding — Refactoring Under Constraints 20Coding — Test Writing 20Coding — React Component (typed) 20Coding — TypeScript Type System 20Coding — React Debugging 20

Doc/OCR vision 1d 160 / 160 20 8/8 n=1 26.3 run 2026-09-24
Test Category Score tok/s
D1 Doc/OCR — Transcription 20/20 17.4
D2 Doc/OCR — Document QA 20/20 16.1
D3 Doc/OCR — Field Extraction 20/20 18.7
D4 Doc/OCR — Document QA 20/20 22.2
D5 Doc/OCR — Transcription 20/20 18.9
D6 Doc/OCR — Field Extraction 20/20 18.7
D7 Doc/OCR — Degraded Input 20/20 22.8
D8 Doc/OCR — Degraded Input 20/20 22.9

Category mean Doc/OCR — Transcription 20Doc/OCR — Document QA 20Doc/OCR — Field Extraction 20Doc/OCR — Degraded Input 20

Doc/OCR — real-degraded tier 1d2 59.2 / 80 14.8 4/4 n=1 28.2 run 2026-09-22
Test Category Score tok/s
D9 Doc/OCR — Real Degraded Transcription 13/20 24.2
D10 Doc/OCR — Handwriting Extraction 16/20 16.9
D11 Doc/OCR — Degraded Table Transcription 14/20 25.8
D12 Doc/OCR — Degraded Table QA 16/20 26.2

Category mean Doc/OCR — Real Degraded Transcription 13Doc/OCR — Handwriting Extraction 16Doc/OCR — Degraded Table Transcription 14Doc/OCR — Degraded Table QA 16

Content-production depth 1e 117 / 120 19.5 6/6 n=1.5 25.6 run 2026-10-04
Test Category Score tok/s
E1 SOW & Requirements Decomposition 20/20 28.4
E2 SOW & Requirements Decomposition 20/20 26.4
E3 Executive & Proposal Content 18/20 27
E4 Executive & Proposal Content 20/20 25.6
E5 Executive & Proposal Content 19/20 24.7
E6 Executive & Proposal Content 20/20 22.2

Category mean SOW & Requirements Decomposition 20Executive & Proposal Content 19.3

Long-context retrieval & synthesis 1g 120 / 120 20 6/6 n=1 22 run 2026-09-24
Test Category Score tok/s
L1 Long-Context — Needle Retrieval 20/20 10
L2 Long-Context — Grounded QA with Distractors 20/20 11.4
L3 Long-Context — Ops Log Reasoning 20/20 8.6
L4 Long-Context — Faithful Summarization 20/20 18.6
L5 Long-Context — Instruction Retention 20/20 9.9
L6 Long-Context — Full-Haystack QA 20/20 3.4

Category mean Long-Context — Needle Retrieval 20Long-Context — Grounded QA with Distractors 20Long-Context — Ops Log Reasoning 20Long-Context — Faithful Summarization 20Long-Context — Instruction Retention 20Long-Context — Full-Haystack QA 20

Long-context multi-needle (MRCR) 1g2 44.1 / 60 14.7 3/3 n=1.3 18 run 2026-09-21
Test Category Score tok/s
L7 Long-Context — Multi-Needle Retrieval (MRCR) 20/20 11
L8 Long-Context — Multi-Needle Retrieval (MRCR) 20/20 7
L9 Long-Context — Multi-Needle Retrieval (MRCR) 4/20 15.7
Live one-shot builds (runtime-verified) 1h 56.1 / 80 18.7 3/4 ● n=1 24 run 2026-09-21
Test Category Score tok/s
H1 Web-Estate One-Shot Builds 17/20 24.1
H2 Web-Estate One-Shot Builds 19/20 23.9
H4 Web-Estate One-Shot Builds 20/20 24.1
Production replay (real agent workload) 1i 178.8 / 240 14.9 12/12 n=1 21.7 run 2026-10-08
Test Category Score tok/s
I1 Nightly Fix Replay 7/20 20.2
I2 Nightly Fix Replay 14/20 20.6
I3 Nightly Fix Replay 9/20 15.9
I4 Nightly Fix Replay 10/20 20.2
I5 Nightly Fix Replay 12/20 20.3
I6 Nightly Fix Replay 13/20 20.5
I7 Audit Judgment 15/20 22.8
I8 Audit Judgment 20/20 22.9
I9 Cron Dependency Reasoning 20/20 21.3
I10 Cron Dependency Reasoning 19/20 20.4
I11 Curation Judgment 20/20 22.3
I12 Curation Judgment 20/20 22.8

Category mean Nightly Fix Replay 10.8Audit Judgment 17.5Cron Dependency Reasoning 19.5Curation Judgment 20

Brand-constrained static build (negative-constraint adherence) 1j 93 / 100 18.6 5/5 n=1 22.8 run 2026-09-24
Test Category Score tok/s
J1 Brand-Constrained Static Build 13/20 22.8
J3 Brand-Constrained Static Build 20/20 23.6
J4 Brand-Constrained Static Build 20/20 23.3
J5 Brand-Constrained Static Build 20/20 22.3
J6 Brand-Constrained Static Build 20/20 21.9

Each test is LLM-judged 0–20 against a fixed rubric; suite max = tests × 20. An ADJ chip marks a cell carrying a subjective adjustment from the maintainer; the judged score is the one shown, is unchanged, and remains what the axes, ranks and tie bands are computed from. Hover for the reason. Rows with a ▶ expand to the per-test breakdown. Where fewer tests ran than the suite total, unrun tests count as zero toward the axis score. See methodology.