Bonsai 2 27B — GGUF vs Qwen3.8-27B vs LFM3-1.7B vs MiniCPM5-2B

New Comparison

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

Winner

Qwen3.8-27B

Qwen3.8-27B scores higher overall across category strength profiles (86 vs 39 for Bonsai 2 27B — GGUF); only 1 benchmarks are directly comparable.

Qwen3.8-27B 85.9 Bonsai 2 27B — GGUF 38.8 LFM3-1.7B 35.8 MiniCPM5-2B 35.2

Only 1 of 88 benchmarks are scored by every selected model — the verdict leans on category strength profiles and is indicative.

Benchmark Scores Comparison

Bonsai 2 27B — GGUF Qwen3.8-27B LFM3-1.7B MiniCPM5-2B

Bonsai 2 27B — GGUF

27.4B params

Qwen3.8-27B

27.0B params

Score: 100.0

LFM3-1.7B

1.7B params

MiniCPM5-2B

2.5B params

Score: 100.0

Coverage matrix

4 models × 88 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 88 benchmarks each model has scores for.
Benchmark Qwen3.8-27B Qwen Team (Alibaba Cloud) MiniCPM5-2B OpenBMB Bonsai 2 27B — GGUF Prism ML LFM3-1.7B LiquidAI
OmniScience Accuracy — — — 7.07
GPQA Diamond 89.2 70.2 — 63.37
IFBench (prompt loose) 79.5 — 74 —
LiveCodeBench v6 90.3 69.1 90.07 73.46
SWE-bench Pro 61.7 14.4 — —
BFCL-V4 72.4 66.6 — 71.39
RealWorldQA 85.9 — — —
Agents' Last Exam (Score) 42.9 — — —
AA-LCR 80.3 59 — 52.83
MathVision 90 — — —
AIME 26 98.8 86.5 95.83 —
BrowseComp-zh — 43.5 — —
Multi-IF — 71.8 — 89.1
AndroidWorld 81.9 — — —
Agents' Last Exam 20.4 — — —
Toolathlon Verified 67.1 — — —
RecreationBench 47.1 — — —
TAU3-Bench — 20.8 — 68.01
SWE-bench Verified (OpenCode) — — — 43.4
OJBench — 32.5 — —
OmniScience Non-Hallucination — — — 92.93
AutomationBench (Public) 58.7 — — 41.29
GSM8K — — 96.66 —
GDPVal 63.5 — — —
RULER 128K — — — 65.56
LCB-Pro 25Q2 (Medium) — 17.5 — —
MBPP+ — — 83.07 —
WildClawBench — 23.9 — —
LCB-Pro 25Q2 (Easy) — 68 — —
NL2Repo 42.3 — — —
MMMU-Pro — — 75.49 —
LiveCodeBench Pro 25Q2 — — — 31.74
QwenClawBench — 42.9 — —
BabyVision 65.7 — — —
Claw-Eval Avg 56.9 59.2 — 68.64
HumanEval+ — — 95.12 —
Multi-Challenge 63.6 — — —
SWE-MM 38.6 — — —
MMLU-Pro 85.7 70.8 — —
SuperGPQA — 40.8 — —
OSWorld-Verified 84.3 — — —
LiveClaw — — — 45.6
Ifstruct V1 — — — 93.72
IFEval — 86.7 91.31 96.66
HLE (with tools) 30.8 — — 13.9
LVBench 72.4 — — —
CharXiv (RQ) 83.7 — — —
GDP.pdf 11.2 — — —
BrowseComp — 39.7 — —
MMLU-Redux — 84.7 89.09 —
NL2Repo-Bench 42.3 — — —
WebArena-Verified 64.8 — — 38.42
CoWorkBench 70.7 — — —
AA-Omniscience Index (public set) -0.5 — — 0.5
SWE-bench Multilingual 73.8 — — 23
NoLiMa — 68.1 — —
OCR Bench v2 — — 56.88 —
IFBench 79.5 66.3 — 73.03
SWE-bench Verified 74.6 46.4 — —
Humanity's Last Exam 30.8 8.9 — 7.17
ERQA 65.5 — — —
QwenSWEBench 79 — — —
TAU2-Bench — 97.1 — —
MATH-500 — 94.6 98.8 —
Terminal-Bench 2.1 (Terminus-2) 73 — — —
OmniDocBench 1.5 91.1 — — —
HMMT Feb 26 95.8 63.8 — —
Terminal-Bench Lite 67 — — —
Terminal Bench 2.1 — 8.6 — 12.92
AIME 2025 — 86.5 95 —
JobBench 33.4 — — —
LongBenchPro — 44.8 — —
MuSR — — 70.63 —
OSWorld 2.0 (Partial) 48 — — —
OSWorld 2.0 (Binary) 19.4 — — —
EnterpriseOps-Gym — — — 28.2
Vision2Web 62.9 — — —
Claw-Eval Pass^3 57.4 — — 51.4
LongBench v2 — 43.7 — —
MCP-Atlas — — — 32.57
DeepSWE 1.1 42.2 — — —
Chatbot Korean (internal) 90.1 — — —
ToolSandbox — — — 80.36
Gaia2 — 88.7 — —
BFCL-V3 — — 74.92 —
SciCode 44.6 26.3 — 28.47
BrowseComp Plus 58 — — 37
GDPVal-AA v2 — 19.6 — —
Covered 50/88 34/88 14/88 29/88

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
Qwen3.8-27B Winner 85.9 1 of 1 50 of 88 (57.0%) Other Instruction Following
Bonsai 2 27B — GGUF 38.8 0 of 1 14 of 88 (16.0%) Math Instruction Following
LFM3-1.7B 35.8 0 of 1 29 of 88 (33.0%) Instruction Following STEM & Reasoning
MiniCPM5-2B 35.2 0 of 1 34 of 88 (39.0%) Long Context General Agents

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 Qwen3.8-27B Bonsai 2 27B — GGUF LFM3-1.7B MiniCPM5-2B
General Agents 26 benchmarks 91 — 62.5 42.2
Agentic Coding 16 benchmarks 89.2 — 65.4 50.3
STEM & Reasoning 11 benchmarks 93.6 73.3 28.2 59.6
Instruction Following 6 benchmarks 77.5 61.6 85.4 66.6
Knowledge 6 benchmarks 84.1 70 70 49.6
Long Context 6 benchmarks 100 — 65.8 73.5
Math 2 benchmarks — 100 — 68.7
Other 1 benchmark 100 — 58.2 52.6

Fully comparable benchmarks

Scored by all 4 selected models (1)

Benchmark Category Bonsai 2 27B — GGUF Qwen3.8-27B LFM3-1.7B MiniCPM5-2B
LiveCodeBench v6 stem_reasoning 90.07 90.3 73.46 69.1

Partial coverage benchmarks

Scored by at least one but not all selected models (87) — missing scores show as —

Benchmark Category Bonsai 2 27B — GGUF Qwen3.8-27B LFM3-1.7B MiniCPM5-2B
OmniScience Accuracy 1/4 models knowledge — — 7.07 —
GPQA Diamond 3/4 models stem_reasoning — 89.2 63.37 70.2
IFBench (prompt loose) 2/4 models instruction_following 74 79.5 — —
SWE-bench Pro 2/4 models coding_agent — 61.7 — 14.4
BFCL-V4 3/4 models general_agent — 72.4 71.39 66.6
RealWorldQA 1/4 models vision_language — 85.9 — —
Agents' Last Exam (Score) 1/4 models general_agent — 42.9 — —
AA-LCR 3/4 models long_context — 80.3 52.83 59
MathVision 1/4 models vision_language — 90 — —
AIME 26 3/4 models stem_reasoning 95.83 98.8 — 86.5
BrowseComp-zh 1/4 models general_agent — — — 43.5
Multi-IF 2/4 models instruction_following — — 89.1 71.8
AndroidWorld 1/4 models general_agent — 81.9 — —
Agents' Last Exam 1/4 models general_agent — 20.4 — —
Toolathlon Verified 1/4 models general_agent — 67.1 — —
RecreationBench 1/4 models general_agent — 47.1 — —
TAU3-Bench 2/4 models general_agent — — 68.01 20.8
SWE-bench Verified (OpenCode) 1/4 models coding_agent — — 43.4 —
OJBench 1/4 models stem_reasoning — — — 32.5
OmniScience Non-Hallucination 1/4 models knowledge — — 92.93 —
AutomationBench (Public) 2/4 models general_agent — 58.7 41.29 —
GSM8K 1/4 models math 96.66 — — —
GDPVal 1/4 models general_agent — 63.5 — —
RULER 128K 1/4 models long_context — — 65.56 —
LCB-Pro 25Q2 (Medium) 1/4 models stem_reasoning — — — 17.5
MBPP+ 1/4 models code 83.07 — — —
WildClawBench 1/4 models coding_agent — — — 23.9
LCB-Pro 25Q2 (Easy) 1/4 models stem_reasoning — — — 68
NL2Repo 1/4 models coding_agent — 42.3 — —
MMMU-Pro 1/4 models vision_language 75.49 — — —
LiveCodeBench Pro 25Q2 1/4 models stem_reasoning — — 31.74 —
QwenClawBench 1/4 models coding_agent — — — 42.9
BabyVision 1/4 models vision_language — 65.7 — —
Claw-Eval Avg 3/4 models coding_agent — 56.9 68.64 59.2
HumanEval+ 1/4 models code 95.12 — — —
Multi-Challenge 1/4 models instruction_following — 63.6 — —
SWE-MM 1/4 models coding_agent — 38.6 — —
MMLU-Pro 2/4 models knowledge — 85.7 — 70.8
SuperGPQA 1/4 models knowledge — — — 40.8
OSWorld-Verified 1/4 models general_agent — 84.3 — —
LiveClaw 1/4 models general_agent — — 45.6 —
Ifstruct V1 1/4 models instruction_following — — 93.72 —
IFEval 3/4 models instruction_following 91.31 — 96.66 86.7
HLE (with tools) 2/4 models stem_reasoning — 30.8 13.9 —
LVBench 1/4 models video_understanding — 72.4 — —
CharXiv (RQ) 1/4 models document_understanding — 83.7 — —
GDP.pdf 1/4 models long_context — 11.2 — —
BrowseComp 1/4 models general_agent — — — 39.7
MMLU-Redux 2/4 models knowledge 89.09 — — 84.7
NL2Repo-Bench 1/4 models coding_agent — 42.3 — —
WebArena-Verified 2/4 models general_agent — 64.8 38.42 —
CoWorkBench 1/4 models general_agent — 70.7 — —
AA-Omniscience Index (public set) 2/4 models knowledge — -0.5 0.5 —
SWE-bench Multilingual 2/4 models coding_agent — 73.8 23 —
NoLiMa 1/4 models long_context — — — 68.1
OCR Bench v2 1/4 models document_understanding 56.88 — — —
IFBench 3/4 models instruction_following — 79.5 73.03 66.3
SWE-bench Verified 2/4 models coding_agent — 74.6 — 46.4
Humanity's Last Exam 3/4 models stem_reasoning — 30.8 7.17 8.9
ERQA 1/4 models spatial_intelligence — 65.5 — —
QwenSWEBench 1/4 models coding_agent — 79 — —
TAU2-Bench 1/4 models general_agent — — — 97.1
MATH-500 2/4 models math 98.8 — — 94.6
Terminal-Bench 2.1 (Terminus-2) 1/4 models coding_agent — 73 — —
OmniDocBench 1.5 1/4 models document_understanding — 91.1 — —
HMMT Feb 26 2/4 models stem_reasoning — 95.8 — 63.8
Terminal-Bench Lite 1/4 models coding_agent — 67 — —
Terminal Bench 2.1 2/4 models coding_agent — — 12.92 8.6
AIME 2025 2/4 models stem_reasoning 95 — — 86.5
JobBench 1/4 models general_agent — 33.4 — —
LongBenchPro 1/4 models long_context — — — 44.8
MuSR 1/4 models reasoning 70.63 — — —
OSWorld 2.0 (Partial) 1/4 models general_agent — 48 — —
OSWorld 2.0 (Binary) 1/4 models general_agent — 19.4 — —
EnterpriseOps-Gym 1/4 models general_agent — — 28.2 —
Vision2Web 1/4 models vision_language — 62.9 — —
Claw-Eval Pass^3 2/4 models coding_agent — 57.4 51.4 —
LongBench v2 1/4 models long_context — — — 43.7
MCP-Atlas 1/4 models general_agent — — 32.57 —
DeepSWE 1.1 1/4 models coding_agent — 42.2 — —
Chatbot Korean (internal) 1/4 models multilingual — 90.1 — —
ToolSandbox 1/4 models general_agent — — 80.36 —
Gaia2 1/4 models general_agent — — — 88.7
BFCL-V3 1/4 models general_agent 74.92 — — —
SciCode 3/4 models — — 44.6 28.47 26.3
BrowseComp Plus 2/4 models general_agent — 58 37 —
GDPVal-AA v2 1/4 models general_agent — — — 19.6

Comparison FAQ

Which model scores highest in this comparison?

Qwen3.8-27B scores higher overall across category strength profiles (86 vs 39 for Bonsai 2 27B — GGUF); only 1 benchmarks are directly comparable.

Which model is strongest at STEM & Reasoning?

Qwen3.8-27B leads the STEM & Reasoning category (strength score 93.6/100 across 7 comparable benchmarks).

How reliable is this comparison verdict?

Only 1 of 88 benchmarks are scored by every selected model — the verdict leans on category strength profiles and is indicative.

What now?

You've seen 4 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