Side-by-side benchmark comparison across agentic, coding, multimodal, knowledge, reasoning, and math workflows.
Llama 4 Behemoth
40
0/8 categoriesQwen3.5-35B-A3B
68
Winner · 7/8 categoriesLlama 4 Behemoth· Qwen3.5-35B-A3B
Pick Qwen3.5-35B-A3B if you want the stronger benchmark profile. Llama 4 Behemoth only becomes the better choice if you would rather avoid the extra latency and token burn of a reasoning model.
Qwen3.5-35B-A3B is clearly ahead on the aggregate, 68 to 40. The gap is large enough that you do not need to squint at the spreadsheet to see the difference.
Qwen3.5-35B-A3B's sharpest advantage is in coding, where it averages 72.6 against 14.2. The single biggest benchmark swing on the page is LiveCodeBench, 13% to 74.6%.
Qwen3.5-35B-A3B is the reasoning model in the pair, while Llama 4 Behemoth is not. That usually helps on harder chain-of-thought-heavy tests, but it can also mean more latency and more token spend in real use. Qwen3.5-35B-A3B gives you the larger context window at 262K, compared with 32K for Llama 4 Behemoth.
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 | Llama 4 Behemoth | Qwen3.5-35B-A3B |
|---|---|---|
| AgenticQwen3.5-35B-A3B wins | ||
| Terminal-Bench 2.0 | 33% | 40.5% |
| BrowseComp | 38% | 61% |
| OSWorld-Verified | 34% | 54.5% |
| tau2-bench | — | 81.2% |
| CodingQwen3.5-35B-A3B wins | ||
| HumanEval | 40% | — |
| SWE-bench Verified | 15% | 69.2% |
| LiveCodeBench | 13% | 74.6% |
| SWE-bench Pro | 15% | — |
| Multimodal & GroundedQwen3.5-35B-A3B wins | ||
| MMMU-Pro | 60% | 75.1% |
| OfficeQA Pro | 49% | — |
| ReasoningQwen3.5-35B-A3B wins | ||
| MuSR | 44% | — |
| BBH | 62% | — |
| LongBench v2 | 46% | 59% |
| MRCRv2 | 46% | — |
| KnowledgeQwen3.5-35B-A3B wins | ||
| MMLU | 48% | — |
| GPQA | 47% | 84.2% |
| SuperGPQA | 45% | 63.4% |
| MMLU-Pro | 54% | 85.3% |
| HLE | 3% | — |
| FrontierScience | 43% | — |
| SimpleQA | 46% | — |
| Instruction FollowingQwen3.5-35B-A3B wins | ||
| IFEval | 68% | 91.9% |
| MultilingualQwen3.5-35B-A3B wins | ||
| MGSM | 66% | — |
| MMLU-ProX | 61% | 81% |
| Mathematics | ||
| AIME 2023 | 48% | — |
| AIME 2024 | 50% | — |
| AIME 2025 | 49% | — |
| HMMT Feb 2023 | 44% | — |
| HMMT Feb 2024 | 46% | — |
| HMMT Feb 2025 | 45% | — |
| BRUMO 2025 | 47% | — |
| MATH-500 | 60% | — |
Qwen3.5-35B-A3B is ahead overall, 68 to 40. The biggest single separator in this matchup is LiveCodeBench, where the scores are 13% and 74.6%.
Qwen3.5-35B-A3B has the edge for knowledge tasks in this comparison, averaging 79.3 versus 37.3. Inside this category, GPQA is the benchmark that creates the most daylight between them.
Qwen3.5-35B-A3B has the edge for coding in this comparison, averaging 72.6 versus 14.2. Inside this category, LiveCodeBench is the benchmark that creates the most daylight between them.
Qwen3.5-35B-A3B has the edge for reasoning in this comparison, averaging 59 versus 45.5. Inside this category, LongBench v2 is the benchmark that creates the most daylight between them.
Qwen3.5-35B-A3B has the edge for agentic tasks in this comparison, averaging 50.5 versus 34.6. Inside this category, BrowseComp is the benchmark that creates the most daylight between them.
Qwen3.5-35B-A3B has the edge for multimodal and grounded tasks in this comparison, averaging 75.1 versus 55.1. Inside this category, MMMU-Pro is the benchmark that creates the most daylight between them.
Qwen3.5-35B-A3B has the edge for instruction following in this comparison, averaging 91.9 versus 68. Inside this category, IFEval is the benchmark that creates the most daylight between them.
Qwen3.5-35B-A3B has the edge for multilingual tasks in this comparison, averaging 81 versus 62.8. Inside this category, MMLU-ProX 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.