Side-by-side benchmark comparison across agentic, coding, multimodal, knowledge, reasoning, and math workflows.
DeepSeek V3.2 (Thinking)
67
Winner · 2/8 categoriesGemma 4 26B A4B
64
2/8 categoriesDeepSeek V3.2 (Thinking)· Gemma 4 26B A4B
Pick DeepSeek V3.2 (Thinking) if you want the stronger benchmark profile. Gemma 4 26B A4B only becomes the better choice if coding is the priority or you need the larger 256K context window.
DeepSeek V3.2 (Thinking) has the cleaner overall profile here, landing at 67 versus 64. It is a real lead, but still close enough that category-level strengths matter more than the headline number.
DeepSeek V3.2 (Thinking)'s sharpest advantage is in reasoning, where it averages 60.1 against 44.1. The single biggest benchmark swing on the page is MRCRv2, 78% to 44.1%. Gemma 4 26B A4B does hit back in coding, so the answer changes if that is the part of the workload you care about most.
Gemma 4 26B A4B gives you the larger context window at 256K, compared with 128K for DeepSeek V3.2 (Thinking).
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 | DeepSeek V3.2 (Thinking) | Gemma 4 26B A4B |
|---|---|---|
| Agentic | ||
| Terminal-Bench 2.0 | 71% | — |
| BrowseComp | 70% | — |
| OSWorld-Verified | 67% | — |
| DeepPlanning | 27.4% | — |
| CodingGemma 4 26B A4B wins | ||
| HumanEval | 79% | — |
| SWE-bench Verified | 48% | — |
| LiveCodeBench | 45% | 77.1% |
| SWE-bench Pro | 58% | — |
| Multimodal & GroundedGemma 4 26B A4B wins | ||
| MMMU-Pro | 66% | 73.8% |
| ReasoningDeepSeek V3.2 (Thinking) wins | ||
| MuSR | 81% | — |
| BBH | 86% | 64.8% |
| LongBench v2 | 78% | — |
| MRCRv2 | 78% | 44.1% |
| ARC-AGI-2 | 4% | — |
| KnowledgeDeepSeek V3.2 (Thinking) wins | ||
| MMLU | 87% | — |
| GPQA | 85% | 82.3% |
| SuperGPQA | 83% | — |
| MMLU-Pro | 73% | 82.6% |
| HLE | 22% | 17.2% |
| FrontierScience | 77% | — |
| SimpleQA | 83% | — |
| HLE w/o tools | — | 8.7% |
| Instruction Following | ||
| IFEval | 85% | — |
| Multilingual | ||
| MGSM | 84% | — |
| MMLU-ProX | 79% | — |
| Mathematics | ||
| AIME 2023 | 87% | — |
| AIME 2024 | 89% | — |
| AIME 2025 | 88% | — |
| HMMT Feb 2023 | 83% | — |
| HMMT Feb 2024 | 85% | — |
| HMMT Feb 2025 | 84% | — |
| BRUMO 2025 | 86% | — |
| MATH-500 | 84% | — |
DeepSeek V3.2 (Thinking) is ahead overall, 67 to 64. The biggest single separator in this matchup is MRCRv2, where the scores are 78% and 44.1%.
DeepSeek V3.2 (Thinking) has the edge for knowledge tasks in this comparison, averaging 65.9 versus 56.1. Inside this category, MMLU-Pro is the benchmark that creates the most daylight between them.
Gemma 4 26B A4B has the edge for coding in this comparison, averaging 77.1 versus 50.7. Inside this category, LiveCodeBench is the benchmark that creates the most daylight between them.
DeepSeek V3.2 (Thinking) has the edge for reasoning in this comparison, averaging 60.1 versus 44.1. Inside this category, MRCRv2 is the benchmark that creates the most daylight between them.
Gemma 4 26B A4B has the edge for multimodal and grounded tasks in this comparison, averaging 73.8 versus 66. Inside this category, MMMU-Pro 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.