Bonsai 2 27B — GGUF vs Inkling 2 Mini vs Qwen3.8-27B

New Comparison

Scores are fetched live from the current database on every visit — nothing is cached.

Winner

Bonsai 2 27B — GGUF

Bonsai 2 27B — GGUF scores higher overall across category strength profiles (68 vs 61 for Qwen3.8-27B); only 1 benchmarks are directly comparable.

Bonsai 2 27B — GGUF 68 Qwen3.8-27B 61.2 Inkling 2 Mini 54.9

Only 1 of 106 benchmarks are scored by every selected model — the verdict leans on category strength profiles and is indicative.

Benchmark Scores Comparison

Bonsai 2 27B — GGUF Inkling 2 Mini Qwen3.8-27B

Bonsai 2 27B — GGUF

27.4B params

Inkling 2 Mini

26.2B params

Qwen3.8-27B

27.0B params

Score: 100.0

Coverage matrix

3 models × 106 benchmarks — hover a benchmark name for its category

Full Coverage Matrix
Hides benchmarks where all selected models score identically.
Shading = score relative to the best score among the selected models: Low Mid High Missing (—) The totals row shows how many of the 106 benchmarks each model has scores for.
Benchmark Qwen3.8-27B Qwen Team (Alibaba Cloud) Bonsai 2 27B — GGUF Prism ML Inkling 2 Mini Thinking Machines
ARC-AGI (v1) 87.5 — 84.5
PerceptionBench 52.4 — 57.5
GPQA Diamond 89.2 — —
IFBench (prompt loose) 79.5 74 —
Vibe Code Bench v1.1 64.9 — 77.9
FrontierCode 2.0 14.9 — 15.1
LiveCodeBench v6 90.3 90.07 —
ComplexConstraints v1.1 30.1 — 33.4
SWE-bench Pro 61.7 — —
BFCL-V4 72.4 — —
RealWorldQA 85.9 — —
Agents' Last Exam (Score) 42.9 — —
AA-LCR 80.3 — —
MathVision 90 — —
Destructive actions (Internal) 35.2 — 10.2
AIME 26 98.8 95.83 —
AndroidWorld 81.9 — —
Agents' Last Exam 20.4 — —
Toolathlon Verified 67.1 — —
RecreationBench 47.1 — —
BVB 70.7 — 68.4
ScreenSpot Pro 57.6 — 76.5
OmniScience Non-Hallucination 71.4 — 63.4
ScreenSpot Pro (with Python) 78.9 — 87.7
AutomationBench (Public) 58.7 — —
ZeroBench_main w/ python (pass@5) — — 34
GSM8K — 96.66 —
MuirBench 65.9 — 77.5
GDPVal 63.5 — —
DeepSearch QA 79.7 — 85.3
StrongREJECT v2 (adaptive) 99 — 97.8
ChartoGraphy w/ python 52.1 — 72.5
MBPP+ — 83.07 —
SWE-bench Pro v2.0 Hard 73.8 — 77.9
NL2Repo 42.3 — —
MMMU-Pro 76.3 75.49 74.9
Chartography 12.2 — 22.1
BabyVision 65.7 — —
Claw-Eval Avg 56.9 — —
HumanEval+ — 95.12 —
Multi-Challenge 63.6 — —
SWE-MM 38.6 — —
Artificial Analysis Intelligence Index v5 20.1 — 20.1
MMLU-Pro 85.7 — —
OSWorld-Verified 84.3 — —
IFEval — 91.31 —
HLE (with tools) 30.8 — —
LVBench 72.4 — —
CharXiv (RQ) 83.7 — —
Automation Bench-AA 48.2 — 41.8
GDP.pdf 11.2 — 16.8
BrowseComp 61.6 — 79
MMLU-Redux — 89.09 —
NL2Repo-Bench 42.3 — —
WebArena-Verified 64.8 — —
ARC-AGI (v2) 42.4 — 49.3
CoWorkBench 70.7 — —
Diligence Bench 46.4 — 48.6
AA-Omniscience Index (public set) -0.5 — —
SWE-bench Multilingual 73.8 — 73.7
CritPt v2 (Full) 10.6 — 31.4
OCR Bench v2 — 56.88 —
PMPP-Hard (AUC, 0-10h) 262.5 — 455.82
IFBench 79.5 — —
SWE-bench Verified 74.6 — 79.2
Humanity's Last Exam 30.8 — —
OmniDocBench v1.6 91.5 — 92.8
ERQA 65.5 — —
HLE-Diamond (no tools) 15.3 — 15.2
Vals Index (v2.2) — — 46.5
QwenSWEBench 79 — —
MATH-500 — 98.8 —
Terminal-Bench 2.1 (Terminus-2) 73 — —
OmniDocBench 1.5 91.1 — —
HMMT Feb 26 95.8 — —
Terminal-Bench Lite 67 — —
MathArena Apex 2025 18.8 — 34.4
Terminal Bench 2.1 73 — 71.6
AIME 2025 — 95 —
JobBench 33.4 — 49.6
CritPt v2 (Hard) 4 — 17
AA-AnalystAgent — — 46.3
MuSR — 70.63 —
Misbehavior (Internal) 57.7 — 9.8
OSWorld 2.0 (Partial) 48 — —
OSWorld 2.0 (Binary) 19.4 — —
EnterpriseOps-Gym 44.2 — 42
Vals Finance Index — — 50.1
Vision2Web 62.9 — —
Claw-Eval Pass^3 57.4 — —
LongBench v2 52.3 — 66.9
MCP-Atlas 77.3 — 82.7
MRCR v2 (8-needle, 256K-512K) 38 — 46.4
Terminal-Bench 4.0 5.6 — 5.6
DeepSWE 1.1 42.2 — 48.2
Chatbot Korean (internal) 90.1 — —
FORTRESS Benign 97.8 — 96.4
AA-Briefcase v1.1 1391 — 1263
BFCL-V3 — 74.92 —
Covert misbehavior (Internal) 33.2 — 8.3
SciCode 44.6 — —
GraphWalks (BFS, 256K) 50.4 — 69.7
BrowseComp Plus 58 — —
AA-Omniscience v1.1 (score) -6.4 — -8.6
GDPval-AA v2.1 1427 — 1319
ZEROBench 12 — 17
Covered 92/106 14/106 51/106

Scorecard

Overall 0–100 score — category strengths (50%), shared-benchmark wins (30%), coverage (20%)

Model Overall score Shared benchmark wins Coverage Strongest at Weakest at
Bonsai 2 27B — GGUF Winner 68 1 of 1 14 of 106 (13.0%) STEM & Reasoning Vision & Language
Qwen3.8-27B 61.2 0 of 1 92 of 106 (87.0%) Safety Multimodal Agents
Inkling 2 Mini 54.9 0 of 1 51 of 106 (48.0%) Video Understanding Safety

Category strengths

Where each model stands, per benchmark category — score relative to the best model, adjusted for benchmark coverage. This also ranks models when they share few benchmarks.

Category Bonsai 2 27B — GGUF Qwen3.8-27B Inkling 2 Mini
General Agents 24 benchmarks — 86.3 97.3
Agentic Coding 16 benchmarks — 99.1 95.7
STEM & Reasoning 9 benchmarks 75.9 88.7 77.2
Vision & Language 9 benchmarks 66.3 80.8 97
Knowledge 5 benchmarks — 100 88.8
Long Context 5 benchmarks — 83.1 92.1
Reasoning 5 benchmarks — 60.8 99.1
Safety 5 benchmarks — 100 53.7
Instruction Following 4 benchmarks 67.6 100 —
Vision 3 benchmarks — 96.2 98.9
Agentic 2 benchmarks — 95.1 93.4
Coding 2 benchmarks — 70.5 100
Composite 2 benchmarks — 100 100
Finance 2 benchmarks — 95.5 100
Video Understanding 2 benchmarks — 85 100
Multimodal Agents 1 benchmark — 55.2 100

Fully comparable benchmarks

Scored by all 3 selected models (1)

Benchmark Category Bonsai 2 27B — GGUF Inkling 2 Mini Qwen3.8-27B
MMMU-Pro vision_language 75.49 74.9 76.3

Partial coverage benchmarks

Scored by at least one but not all selected models (105) — missing scores show as —

Benchmark Category Bonsai 2 27B — GGUF Inkling 2 Mini Qwen3.8-27B
ARC-AGI (v1) 2/3 models reasoning — 84.5 87.5
PerceptionBench 2/3 models vision_language — 57.5 52.4
GPQA Diamond 1/3 models stem_reasoning — — 89.2
IFBench (prompt loose) 2/3 models instruction_following 74 — 79.5
Vibe Code Bench v1.1 2/3 models coding — 77.9 64.9
FrontierCode 2.0 2/3 models coding_agent — 15.1 14.9
LiveCodeBench v6 2/3 models stem_reasoning 90.07 — 90.3
ComplexConstraints v1.1 2/3 models agentic — 33.4 30.1
SWE-bench Pro 1/3 models coding_agent — — 61.7
BFCL-V4 1/3 models general_agent — — 72.4
RealWorldQA 1/3 models vision_language — — 85.9
Agents' Last Exam (Score) 1/3 models general_agent — — 42.9
AA-LCR 1/3 models long_context — — 80.3
MathVision 1/3 models vision_language — — 90
Destructive actions (Internal) 2/3 models safety — 10.2 35.2
AIME 26 2/3 models stem_reasoning 95.83 — 98.8
AndroidWorld 1/3 models general_agent — — 81.9
Agents' Last Exam 1/3 models general_agent — — 20.4
Toolathlon Verified 1/3 models general_agent — — 67.1
RecreationBench 1/3 models general_agent — — 47.1
BVB 2/3 models vision — 68.4 70.7
ScreenSpot Pro 2/3 models general_agent — 76.5 57.6
OmniScience Non-Hallucination 2/3 models knowledge — 63.4 71.4
ScreenSpot Pro (with Python) 2/3 models vision — 87.7 78.9
AutomationBench (Public) 1/3 models general_agent — — 58.7
ZeroBench_main w/ python (pass@5) 1/3 models vision_language — 34 —
GSM8K 1/3 models math 96.66 — —
MuirBench 2/3 models video_understanding — 77.5 65.9
GDPVal 1/3 models general_agent — — 63.5
DeepSearch QA 2/3 models general_agent — 85.3 79.7
StrongREJECT v2 (adaptive) 2/3 models safety — 97.8 99
ChartoGraphy w/ python 2/3 models vision_language — 72.5 52.1
MBPP+ 1/3 models code 83.07 — —
SWE-bench Pro v2.0 Hard 2/3 models coding_agent — 77.9 73.8
NL2Repo 1/3 models coding_agent — — 42.3
Chartography 2/3 models multimodal_agent — 22.1 12.2
BabyVision 1/3 models vision_language — — 65.7
Claw-Eval Avg 1/3 models coding_agent — — 56.9
HumanEval+ 1/3 models code 95.12 — —
Multi-Challenge 1/3 models instruction_following — — 63.6
SWE-MM 1/3 models coding_agent — — 38.6
Artificial Analysis Intelligence Index v5 2/3 models composite — 20.1 20.1
MMLU-Pro 1/3 models knowledge — — 85.7
OSWorld-Verified 1/3 models general_agent — — 84.3
IFEval 1/3 models instruction_following 91.31 — —
HLE (with tools) 1/3 models stem_reasoning — — 30.8
LVBench 1/3 models video_understanding — — 72.4
CharXiv (RQ) 1/3 models document_understanding — — 83.7
Automation Bench-AA 2/3 models agentic — 41.8 48.2
GDP.pdf 2/3 models long_context — 16.8 11.2
BrowseComp 2/3 models general_agent — 79 61.6
MMLU-Redux 1/3 models knowledge 89.09 — —
NL2Repo-Bench 1/3 models coding_agent — — 42.3
WebArena-Verified 1/3 models general_agent — — 64.8
ARC-AGI (v2) 2/3 models reasoning — 49.3 42.4
CoWorkBench 1/3 models general_agent — — 70.7
Diligence Bench 2/3 models finance — 48.6 46.4
AA-Omniscience Index (public set) 1/3 models knowledge — — -0.5
SWE-bench Multilingual 2/3 models coding_agent — 73.7 73.8
CritPt v2 (Full) 2/3 models reasoning — 31.4 10.6
OCR Bench v2 1/3 models document_understanding 56.88 — —
PMPP-Hard (AUC, 0-10h) 2/3 models coding — 455.82 262.5
IFBench 1/3 models instruction_following — — 79.5
SWE-bench Verified 2/3 models coding_agent — 79.2 74.6
Humanity's Last Exam 1/3 models stem_reasoning — — 30.8
OmniDocBench v1.6 2/3 models vision — 92.8 91.5
ERQA 1/3 models spatial_intelligence — — 65.5
HLE-Diamond (no tools) 2/3 models stem_reasoning — 15.2 15.3
Vals Index (v2.2) 1/3 models composite — 46.5 —
QwenSWEBench 1/3 models coding_agent — — 79
MATH-500 1/3 models math 98.8 — —
Terminal-Bench 2.1 (Terminus-2) 1/3 models coding_agent — — 73
OmniDocBench 1.5 1/3 models document_understanding — — 91.1
HMMT Feb 26 1/3 models stem_reasoning — — 95.8
Terminal-Bench Lite 1/3 models coding_agent — — 67
MathArena Apex 2025 2/3 models stem_reasoning — 34.4 18.8
Terminal Bench 2.1 2/3 models coding_agent — 71.6 73
AIME 2025 1/3 models stem_reasoning 95 — —
JobBench 2/3 models general_agent — 49.6 33.4
CritPt v2 (Hard) 2/3 models reasoning — 17 4
AA-AnalystAgent 1/3 models general_agent — 46.3 —
MuSR 1/3 models reasoning 70.63 — —
Misbehavior (Internal) 2/3 models safety — 9.8 57.7
OSWorld 2.0 (Partial) 1/3 models general_agent — — 48
OSWorld 2.0 (Binary) 1/3 models general_agent — — 19.4
EnterpriseOps-Gym 2/3 models general_agent — 42 44.2
Vals Finance Index 1/3 models finance — 50.1 —
Vision2Web 1/3 models vision_language — — 62.9
Claw-Eval Pass^3 1/3 models coding_agent — — 57.4
LongBench v2 2/3 models long_context — 66.9 52.3
MCP-Atlas 2/3 models general_agent — 82.7 77.3
MRCR v2 (8-needle, 256K-512K) 2/3 models long_context — 46.4 38
Terminal-Bench 4.0 2/3 models coding_agent — 5.6 5.6
DeepSWE 1.1 2/3 models coding_agent — 48.2 42.2
Chatbot Korean (internal) 1/3 models multilingual — — 90.1
FORTRESS Benign 2/3 models safety — 96.4 97.8
AA-Briefcase v1.1 2/3 models general_agent — 1263 1391
BFCL-V3 1/3 models general_agent 74.92 — —
Covert misbehavior (Internal) 2/3 models safety — 8.3 33.2
SciCode 1/3 models — — — 44.6
GraphWalks (BFS, 256K) 2/3 models long_context — 69.7 50.4
BrowseComp Plus 1/3 models general_agent — — 58
AA-Omniscience v1.1 (score) 2/3 models knowledge — -8.6 -6.4
GDPval-AA v2.1 2/3 models general_agent — 1319 1427
ZEROBench 2/3 models vision_language — 17 12

Comparison FAQ

Which model scores highest in this comparison?

Bonsai 2 27B — GGUF scores higher overall across category strength profiles (68 vs 61 for Qwen3.8-27B); only 1 benchmarks are directly comparable.

Which model is strongest at General Agents?

Inkling 2 Mini leads the General Agents category (strength score 97.3/100 across 8 comparable benchmarks).

How reliable is this comparison verdict?

Only 1 of 106 benchmarks are scored by every selected model — the verdict leans on category strength profiles and is indicative.

What now?

You've seen 3 models side by side — keep the loop going.

Share this comparison

The URL is the comparison — copy or share it anywhere. No tracking, no cookies.

Not the right matchup?

Open the create form with these models pre-filled and untick the one you want to replace.

Swap a model