Side-by-side benchmark comparison across agentic, coding, multimodal, knowledge, reasoning, and math workflows.
DeepSeek V3.1 (Reasoning)
43
Winner · 1/8 categoriesLFM2.5-350M
~39
1/8 categoriesDeepSeek V3.1 (Reasoning)· LFM2.5-350M
Pick DeepSeek V3.1 (Reasoning) if you want the stronger benchmark profile. LFM2.5-350M only becomes the better choice if instruction following is the priority or you would rather avoid the extra latency and token burn of a reasoning model.
DeepSeek V3.1 (Reasoning) is clearly ahead on the aggregate, 43 to 39. The gap is large enough that you do not need to squint at the spreadsheet to see the difference.
DeepSeek V3.1 (Reasoning)'s sharpest advantage is in knowledge, where it averages 32.5 against 23.8. The single biggest benchmark swing on the page is MMLU-Pro, 53% to 20.0%. LFM2.5-350M does hit back in instruction following, so the answer changes if that is the part of the workload you care about most.
DeepSeek V3.1 (Reasoning) is the reasoning model in the pair, while LFM2.5-350M 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. DeepSeek V3.1 (Reasoning) gives you the larger context window at 128K, compared with 32K for LFM2.5-350M.
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 V3.1 (Reasoning) | LFM2.5-350M |
|---|---|---|
| Agentic | ||
| Terminal-Bench 2.0 | 42% | — |
| BrowseComp | 48% | — |
| OSWorld-Verified | 44% | — |
| Coding | ||
| HumanEval | 26% | — |
| SWE-bench Verified | 14% | — |
| LiveCodeBench | 16% | — |
| SWE-bench Pro | 25% | — |
| Multimodal & Grounded | ||
| MMMU-Pro | 37% | — |
| OfficeQA Pro | 47% | — |
| Reasoning | ||
| MuSR | 30% | — |
| BBH | 64% | — |
| LongBench v2 | 57% | — |
| MRCRv2 | 56% | — |
| KnowledgeDeepSeek V3.1 (Reasoning) wins | ||
| MMLU | 34% | — |
| GPQA | 33% | 30.6% |
| SuperGPQA | 31% | — |
| MMLU-Pro | 53% | 20.0% |
| HLE | 10% | — |
| FrontierScience | 37% | — |
| SimpleQA | 32% | — |
| Instruction FollowingLFM2.5-350M wins | ||
| IFEval | 70% | 77.0% |
| Multilingual | ||
| MGSM | 64% | — |
| MMLU-ProX | 61% | — |
| Mathematics | ||
| AIME 2023 | 34% | — |
| AIME 2024 | 36% | — |
| AIME 2025 | 35% | — |
| HMMT Feb 2023 | 30% | — |
| HMMT Feb 2024 | 32% | — |
| HMMT Feb 2025 | 31% | — |
| BRUMO 2025 | 33% | — |
| MATH-500 | 62% | — |
DeepSeek V3.1 (Reasoning) is ahead overall, 43 to 39. The biggest single separator in this matchup is MMLU-Pro, where the scores are 53% and 20.0%.
DeepSeek V3.1 (Reasoning) has the edge for knowledge tasks in this comparison, averaging 32.5 versus 23.8. Inside this category, MMLU-Pro is the benchmark that creates the most daylight between them.
LFM2.5-350M has the edge for instruction following in this comparison, averaging 77 versus 70. 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.