geekradar
Int
Explore GeekRadar

More to explore

Good gear All topics About us Say hello
GeekRadar / The comparison desk
COMPARE.
AI models
alibaba

Qwen3.6 27B

AI model Released

Context window
262,144 tokens
Maximum output
65,536 tokens
Open weights
Yes

Capability & preference

General capability · ECI

A cross-benchmark capability estimate. Index points, not a percentage; intervals show uncertainty.

Qwen3.6 27B 146.7* Qwen3.6 27B · 95% interval 144.5–148.9

Model performance

Radar task score: four tests, equal weight. An experimental summary of these tasks, not a universal intelligence rating.

Qwen3.6 27B

Radar task score

Unrated

3/4 core tests · not enough evidence

4 benchmark results for this configuration

Scientific reasoning

GPQA Diamond · 25%

Expert-written science questions; strict multiple-choice scoring. Chance accuracy is 25%.

Qwen3.6 27B 84.8*% Approx. 95% interval: 79.8–89.9%

Novel problem solving

Mystery Game Puzzles · 25%

Best-next-move puzzles in an undisclosed game variant. A narrow reasoning test, not general intelligence.

Qwen3.6 27B 7.0*% Approx. 95% interval: 2.0–12.0%

Advanced mathematics

FrontierMath Tiers 1-3 · 25%

Private v2 mathematics problems with Python tools. Earlier versions and Tier 4 are excluded.

Qwen3.6 27B 34.0*% Approx. 95% interval: 28.5–39.5%

Factual answers

SimpleQA Verified · 25%

Correct factual answers under a guessing prompt, graded by a model. Not a hallucination or abstention score.

Qwen3.6 27B Not available

Competition mathematics

OTIS Mock AIME 2024-2025 · Separate result

Competition problems with model-extracted answers checked against an exact key. Earlier grading procedures are excluded.

Qwen3.6 27B 66.7*% Approx. 95% interval: 52.7–80.6%

Bars show reported task scores. Whiskers are approximate 95% intervals where standard errors are available. Different reasoning budgets are not equal-cost tests. Missing results are never zero.

Benchmark results

  • Agentic coding DeepSWE 1.1 · 1.1 · reported score · Claude Code13.30*Platform unreported Compare results
    Test conditions
    Metric
    reported score*
    Checkpoint version
    Not specified*
    Harness
    Claude Code*
    Temperature
    1.0*
    Top p
    0.95*
    Context length
    256K tokens*
  • Agentic coding SWE-bench Pro · version unreported · reported score · Claude Code · Refined task set with corrected problematic tasks53.50*Platform unreported Compare results
    Test conditions
    Metric
    reported score*
    Checkpoint version
    Not specified*
    Harness
    Claude Code*
    Task set
    Refined task set with corrected problematic tasks*
    Temperature
    1.0*
    Top p
    0.95*
    Context length
    256K tokens*
  • Agentic terminal coding Terminal Bench 2.1 (Terminus) · 2.1 · reported score · Terminus63.40*Platform unreported Compare results
    Test conditions
    Metric
    reported score*
    Checkpoint version
    Not specified*
    Harness
    Terminus*
  • Application recreation RecreationBench · version unreported · reported score · Internal29.80*Platform unreported Compare results
    Test conditions
    Metric
    reported score*
    Checkpoint version
    Not specified*
    Task set
    Internal*
    Variant
    Application recreation across Ubuntu, macOS, Windows, Android and web*
  • Browser use WebArena-Verified · version unreported · reported score · OSWorld scaffold; official WebArena-Verified grader48.80*Platform unreported Compare results
    Test conditions
    Metric
    reported score*
    Checkpoint version
    Not specified*
    Harness
    OSWorld scaffold; official WebArena-Verified grader*
  • Competitive coding LiveCodeBench v6 · 6 · reported score83.90*Platform unreported Compare results
    Test conditions
    Metric
    reported score*
    Checkpoint version
    Not specified*
  • Computer use OSWorld-Verified · version unreported · reported score63.90*Platform unreported Compare results
    Test conditions
    Metric
    reported score*
    Checkpoint version
    Not specified*
  • Document intelligence OmniDocBench 1.5 · 1.5 · reported score89.40*Platform unreported Compare results
    Test conditions
    Metric
    reported score*
    Checkpoint version
    Not specified*
  • Embodied intelligence ERQA · version unreported · reported score62.50*Platform unreported Compare results
    Test conditions
    Metric
    reported score*
    Checkpoint version
    Not specified*
  • Frontier agentic tasks Agents' Last Exam · version unreported · pass@110.60*Platform unreported Compare results
    Test conditions
    Metric
    pass@1*
    Checkpoint version
    Not specified*
  • Frontier agentic tasks Agents' Last Exam · version unreported · score27.30*Platform unreported Compare results
    Test conditions
    Metric
    score*
    Checkpoint version
    Not specified*
  • General visual reasoning BabyVision · version unreported · reported score · Code interpreter disabled28.90*Platform unreported Compare results
    Test conditions
    Metric
    reported score*
    Checkpoint version
    Not specified*
    Tool access
    Code interpreter disabled*
  • Instruction following IFBench · version unreported · reported score69.10*Platform unreported Compare results
    Test conditions
    Metric
    reported score*
    Checkpoint version
    Not specified*
  • Long-horizon office work CoWorkBench · version unreported · reported score · Internal61*Platform unreported Compare results
    Test conditions
    Metric
    reported score*
    Checkpoint version
    Not specified*
    Task set
    Internal*
  • Mobile use AndroidWorld · version unreported · reported score70.30*Platform unreported Compare results
    Test conditions
    Metric
    reported score*
    Checkpoint version
    Not specified*
  • Multidisciplinary reasoning HLE · version unreported · reported score24*Platform unreported Compare results
    Test conditions
    Metric
    reported score*
    Checkpoint version
    Not specified*
    Judge
    GPT-4o*
  • Multimodal software engineering SWE-MM · version unreported · reported score · Claude Code · Public SWE-bench Multimodal dev split with Opus 4.7 system-card appendix 8.3 modifications25.70*Platform unreported Compare results
    Test conditions
    Metric
    reported score*
    Checkpoint version
    Not specified*
    Harness
    Claude Code*
    Task set
    Public SWE-bench Multimodal dev split with Opus 4.7 system-card appendix 8.3 modifications*
  • Multimodal tool use ClawEval-MM · version unreported · average score · Mean benchmark score over three trials50.40*Platform unreported Compare results
    Test conditions
    Metric
    average score*
    Checkpoint version
    Not specified*
    Statistic
    Mean benchmark score over three trials*
  • Multimodal tool use ClawEval-MM · version unreported · pass@3 · At least one success in three trials42.60* %Platform unreported Compare results
    Test conditions
    Metric
    pass@3*
    Checkpoint version
    Not specified*
    Statistic
    At least one success in three trials*
  • Professional job tasks JobBench · version unreported · reported score21.80*Platform unreported Compare results
    Test conditions
    Metric
    reported score*
    Checkpoint version
    Not specified*
  • Real-world perception RealWorldQA · version unreported · reported score84.10*Platform unreported Compare results
    Test conditions
    Metric
    reported score*
    Checkpoint version
    Not specified*
  • Repo-level code generation NL2Repo-Bench · version unreported · reported score · Claude Code · Repository acquisition commands disabled36.20*Platform unreported Compare results
    Test conditions
    Metric
    reported score*
    Checkpoint version
    Not specified*
    Harness
    Claude Code*
    Tool access
    Repository acquisition commands disabled*
  • Scientific chart analysis CharXiv (RQ) · version unreported · reported score · Code interpreter disabled78.40*Platform unreported Compare results
    Test conditions
    Metric
    reported score*
    Checkpoint version
    Not specified*
    Tool access
    Code interpreter disabled*
    Annotation set
    Corrected ground-truth annotations*
  • Scientific reasoning GPQA Diamond · version unreported · reported score87.80*Platform unreported Compare results
    Test conditions
    Metric
    reported score*
    Checkpoint version
    Not specified*
  • Software engineering QwenSWEBench · version unreported · reported score · Claude Code · Internal · avg@349.30*Platform unreported Compare results
    Test conditions
    Metric
    reported score*
    Checkpoint version
    Not specified*
    Harness
    Claude Code*
    Statistic
    avg@3*
    Timeout
    8 hours*
    Task set
    Internal*
    Maximum output
    32,768*
    Temperature
    1.0*
    Context length
    256K tokens*
  • Visual math problem solving MathVision · version unreported · reported score · Code interpreter disabled85.10*Platform unreported Compare results
    Test conditions
    Metric
    reported score*
    Checkpoint version
    Not specified*
    Tool access
    Code interpreter disabled*
    Annotation set
    Corrected ground-truth annotations*
    Prompt
    Best of two prompts, with and without boxed final-answer formatting*
  • Visual web development Vision2Web · version unreported · reported score · Claude Code · Category average45*Platform unreported Compare results
    Test conditions
    Metric
    reported score*
    Checkpoint version
    Not specified*
    Harness
    Claude Code*
    Judge
    gpt-5.4-2026-03-05*
    Categories
    Frontend, webpage and website*
    Statistic
    Category average*

Run it locally

Inference · memory estimate

Qwen3.6 27B

51.7 GiB checkpoint tensor payload · 27.78 B stored tensor elements · BF16

Stored tensors can include buffers and repeated shared weights. A verified unique parameter count is unavailable. Checkpoint size is not the memory required to run the model.

API cost

USD / month

Standard text-token estimate. Excludes taxes, caching, batch discounts, reasoning and multimodal surcharges. Provider limits and regional availability may differ.

Qwen3.6 27B

Qwen3.6 27B / Pricing tiers 2
Provider / scopeRateCurrencyAmount / unit
DeepInfraQwen/Qwen3.6-27BStandard InputUSD0.32 / 1M tokens
OutputUSD3.2 / 1M tokens
SiliconFlowQwen/Qwen3.6-27BStandard InputCNY3 / 1M tokens
OutputCNY18 / 1M tokens

Capabilities

Context window
262,144 tokens
Maximum output
65,536 tokens
Reasoning
Yes
Tool calling
Yes
Open weights
Yes
Model license
apache-2.0
Checkpoint tensor payload
51.75 GiB

Input modalities

  • text
  • image
  • video

Output modalities

  • text

Processing & platform

Architecture
qwen3_5

System configuration

Detailed specifications

Specifications · 27 fields
Frontier agentic tasks Agents' Last Exam
suite
Frontier agentic tasks Agents' Last Exam
workload
pass@1
score
10.6*
Scope and conditions
Test conditions
Metric
pass@1*
Checkpoint version
Not specified*
Visual math problem solving MathVision
suite
Visual math problem solving MathVision
workload
reported score · Code interpreter disabled
score
85.1*
Scope and conditions
Test conditions
Metric
reported score*
Checkpoint version
Not specified*
Tool access
Code interpreter disabled*
Annotation set
Corrected ground-truth annotations*
Prompt
Best of two prompts, with and without boxed final-answer formatting*
Frontier agentic tasks Agents' Last Exam
suite
Frontier agentic tasks Agents' Last Exam
workload
score
score
27.3*
Scope and conditions
Test conditions
Metric
score*
Checkpoint version
Not specified*
General visual reasoning BabyVision
suite
General visual reasoning BabyVision
workload
reported score · Code interpreter disabled
score
28.9*
Scope and conditions
Test conditions
Metric
reported score*
Checkpoint version
Not specified*
Tool access
Code interpreter disabled*
Scientific chart analysis CharXiv (RQ)
suite
Scientific chart analysis CharXiv (RQ)
workload
reported score · Code interpreter disabled
score
78.4*
Scope and conditions
Test conditions
Metric
reported score*
Checkpoint version
Not specified*
Tool access
Code interpreter disabled*
Annotation set
Corrected ground-truth annotations*
Instruction following IFBench
suite
Instruction following IFBench
workload
reported score
score
69.1*
Scope and conditions
Test conditions
Metric
reported score*
Checkpoint version
Not specified*
Document intelligence OmniDocBench 1.5
suite
Document intelligence OmniDocBench 1.5
version
1.5
workload
reported score
score
89.4*
Scope and conditions
Test conditions
Metric
reported score*
Checkpoint version
Not specified*
Scientific reasoning GPQA Diamond
suite
Scientific reasoning GPQA Diamond
workload
reported score
score
87.8*
Scope and conditions
Test conditions
Metric
reported score*
Checkpoint version
Not specified*
Multidisciplinary reasoning HLE
suite
Multidisciplinary reasoning HLE
workload
reported score
score
24.0*
Scope and conditions
Test conditions
Metric
reported score*
Checkpoint version
Not specified*
Judge
GPT-4o*
Real-world perception RealWorldQA
suite
Real-world perception RealWorldQA
workload
reported score
score
84.1*
Scope and conditions
Test conditions
Metric
reported score*
Checkpoint version
Not specified*
Embodied intelligence ERQA
suite
Embodied intelligence ERQA
workload
reported score
score
62.5*
Scope and conditions
Test conditions
Metric
reported score*
Checkpoint version
Not specified*
Competitive coding LiveCodeBench v6
suite
Competitive coding LiveCodeBench v6
version
6
workload
reported score
score
83.9*
Scope and conditions
Test conditions
Metric
reported score*
Checkpoint version
Not specified*
Agentic terminal coding Terminal Bench 2.1 (Terminus)
suite
Agentic terminal coding Terminal Bench 2.1 (Terminus)
version
2.1
workload
reported score · Terminus
score
63.4*
Scope and conditions
Test conditions
Metric
reported score*
Checkpoint version
Not specified*
Harness
Terminus*
Computer use OSWorld-Verified
suite
Computer use OSWorld-Verified
workload
reported score
score
63.9*
Scope and conditions
Test conditions
Metric
reported score*
Checkpoint version
Not specified*
Agentic coding SWE-bench Pro
suite
Agentic coding SWE-bench Pro
workload
reported score · Claude Code · Refined task set with corrected problematic tasks
score
53.5*
Scope and conditions
Test conditions
Metric
reported score*
Checkpoint version
Not specified*
Harness
Claude Code*
Task set
Refined task set with corrected problematic tasks*
Temperature
1.0*
Top p
0.95*
Context length
256K tokens*
Browser use WebArena-Verified
suite
Browser use WebArena-Verified
workload
reported score · OSWorld scaffold; official WebArena-Verified grader
score
48.8*
Scope and conditions
Test conditions
Metric
reported score*
Checkpoint version
Not specified*
Harness
OSWorld scaffold; official WebArena-Verified grader*
Mobile use AndroidWorld
suite
Mobile use AndroidWorld
workload
reported score
score
70.3*
Scope and conditions
Test conditions
Metric
reported score*
Checkpoint version
Not specified*
Repo-level code generation NL2Repo-Bench
suite
Repo-level code generation NL2Repo-Bench
workload
reported score · Claude Code · Repository acquisition commands disabled
score
36.2*
Scope and conditions
Test conditions
Metric
reported score*
Checkpoint version
Not specified*
Harness
Claude Code*
Tool access
Repository acquisition commands disabled*
Agentic coding DeepSWE 1.1
suite
Agentic coding DeepSWE 1.1
version
1.1
workload
reported score · Claude Code
score
13.3*
Scope and conditions
Test conditions
Metric
reported score*
Checkpoint version
Not specified*
Harness
Claude Code*
Temperature
1.0*
Top p
0.95*
Context length
256K tokens*
Application recreation RecreationBench
suite
Application recreation RecreationBench
workload
reported score · Internal
score
29.8*
Scope and conditions
Test conditions
Metric
reported score*
Checkpoint version
Not specified*
Task set
Internal*
Variant
Application recreation across Ubuntu, macOS, Windows, Android and web*
Software engineering QwenSWEBench
suite
Software engineering QwenSWEBench
workload
reported score · Claude Code · Internal · avg@3
score
49.3*
Scope and conditions
Test conditions
Metric
reported score*
Checkpoint version
Not specified*
Harness
Claude Code*
Statistic
avg@3*
Timeout
8 hours*
Task set
Internal*
Maximum output
32,768*
Temperature
1.0*
Context length
256K tokens*
Multimodal tool use ClawEval-MM
suite
Multimodal tool use ClawEval-MM
workload
pass@3 · At least one success in three trials
score
42.6*
unit
%
Scope and conditions
Test conditions
Metric
pass@3*
Checkpoint version
Not specified*
Statistic
At least one success in three trials*
Multimodal tool use ClawEval-MM
suite
Multimodal tool use ClawEval-MM
workload
average score · Mean benchmark score over three trials
score
50.4*
Scope and conditions
Test conditions
Metric
average score*
Checkpoint version
Not specified*
Statistic
Mean benchmark score over three trials*
Multimodal software engineering SWE-MM
suite
Multimodal software engineering SWE-MM
workload
reported score · Claude Code · Public SWE-bench Multimodal dev split with Opus 4.7 system-card appendix 8.3 modifications
score
25.7*
Scope and conditions
Test conditions
Metric
reported score*
Checkpoint version
Not specified*
Harness
Claude Code*
Task set
Public SWE-bench Multimodal dev split with Opus 4.7 system-card appendix 8.3 modifications*
Long-horizon office work CoWorkBench
suite
Long-horizon office work CoWorkBench
workload
reported score · Internal
score
61.0*
Scope and conditions
Test conditions
Metric
reported score*
Checkpoint version
Not specified*
Task set
Internal*
Visual web development Vision2Web
suite
Visual web development Vision2Web
workload
reported score · Claude Code · Category average
score
45.0*
Scope and conditions
Test conditions
Metric
reported score*
Checkpoint version
Not specified*
Harness
Claude Code*
Judge
gpt-5.4-2026-03-05*
Categories
Frontend, webpage and website*
Statistic
Category average*
Professional job tasks JobBench
suite
Professional job tasks JobBench
workload
reported score
score
21.8*
Scope and conditions
Test conditions
Metric
reported score*
Checkpoint version
Not specified*

* Provisional data