Side-by-side benchmark comparison across agentic, coding, multimodal, knowledge, reasoning, and math workflows.
1-bit Bonsai 4B
~44
0/8 categoriesDeepSeek V3.2
62
Winner · 4/8 categories1-bit Bonsai 4B· DeepSeek V3.2
Pick DeepSeek V3.2 if you want the stronger benchmark profile. 1-bit Bonsai 4B only becomes the better choice if its workflow or ecosystem matters more than the raw scoreboard.
DeepSeek V3.2 is clearly ahead on the aggregate, 62 to 44. The gap is large enough that you do not need to squint at the spreadsheet to see the difference.
DeepSeek V3.2's sharpest advantage is in knowledge, where it averages 61.8 against 28.7. The single biggest benchmark swing on the page is GPQA, 28.7% to 83%.
DeepSeek V3.2 gives you the larger context window at 128K, compared with 32K for 1-bit Bonsai 4B.
BenchLM keeps the benchmark table and the operator tradeoffs on the same page so a better score does not hide a materially slower, pricier, or smaller-context model.
Runtime metrics show N/A when BenchLM does not have a sourced snapshot for that exact model. The scoring rules and freshness policy are documented on the methodology page.
| Benchmark | 1-bit Bonsai 4B | DeepSeek V3.2 |
|---|---|---|
| Agentic | ||
| Terminal-Bench 2.0 | — | 60% |
| BrowseComp | — | 62% |
| OSWorld-Verified | — | 55% |
| Coding | ||
| HumanEval | — | 76% |
| SWE-bench Verified | — | 45% |
| LiveCodeBench | — | 39% |
| SWE-bench Pro | — | 47% |
| SWE-Rebench | — | 60.9% |
| React Native Evals | — | 69% |
| Multimodal & Grounded | ||
| MMMU-Pro | — | 61% |
| OfficeQA Pro | — | 72% |
| ReasoningDeepSeek V3.2 wins | ||
| MuSR | 41.4% | 79% |
| BBH | — | 81% |
| LongBench v2 | — | 69% |
| MRCRv2 | — | 70% |
| ARC-AGI-2 | — | 4% |
| KnowledgeDeepSeek V3.2 wins | ||
| GPQA | 28.7% | 83% |
| MMLU | — | 84% |
| SuperGPQA | — | 81% |
| MMLU-Pro | — | 73% |
| HLE | — | 11% |
| FrontierScience | — | 72% |
| SimpleQA | — | 81% |
| Instruction FollowingDeepSeek V3.2 wins | ||
| IFEval | 69.6% | 85% |
| Multilingual | ||
| MGSM | — | 84% |
| MMLU-ProX | — | 81% |
| MathematicsDeepSeek V3.2 wins | ||
| MATH-500 | 65.8% | 81% |
| AIME 2023 | — | 84% |
| AIME 2024 | — | 86% |
| AIME 2025 | — | 85% |
| HMMT Feb 2023 | — | 80% |
| HMMT Feb 2024 | — | 82% |
| HMMT Feb 2025 | — | 81% |
| BRUMO 2025 | — | 83% |
DeepSeek V3.2 is ahead overall, 62 to 44. The biggest single separator in this matchup is GPQA, where the scores are 28.7% and 83%.
DeepSeek V3.2 has the edge for knowledge tasks in this comparison, averaging 61.8 versus 28.7. Inside this category, GPQA is the benchmark that creates the most daylight between them.
DeepSeek V3.2 has the edge for math in this comparison, averaging 83.3 versus 65.8. Inside this category, MATH-500 is the benchmark that creates the most daylight between them.
DeepSeek V3.2 has the edge for reasoning in this comparison, averaging 55 versus 41.4. Inside this category, MuSR is the benchmark that creates the most daylight between them.
DeepSeek V3.2 has the edge for instruction following in this comparison, averaging 85 versus 69.6. Inside this category, IFEval is the benchmark that creates the most daylight between them.
Get notified when new models drop, benchmark scores change, or the leaderboard shifts. One email per week.
Free. No spam. Unsubscribe anytime. We only store derived location metadata for consent routing.