LLM benchmarks leaderboard
A side-by-side look at three leading open-weight model families: DeepSeek's reasoning flagship, Alibaba's Qwen3, and Meta's Llama 3.1 405B. Every score below is the model creator's own published number, cited to its source, on knowledge, reasoning, math and coding benchmarks.
At-a-glance comparison
| Model | Parameters | Knowledge | Reasoning | Math | Coding | Source |
|---|---|---|---|---|---|---|
| DeepSeek-R1 Reasoning | 671B total / 37B active | 90.8 MMLU | 71.5 GPQA-Diamond | 79.8 AIME 2024 | 65.9 LiveCodeBench | DeepSeek-R1 paper, Table 4 |
| Qwen3-235B-A22B Thinking mode | 235B total / 22B active | 92.7 MMLU-Redux | 71.1 GPQA-Diamond | 85.7 AIME 2024 | 70.7 LiveCodeBench v5 | Qwen3 Technical Report, Table 11 |
| Llama 3.1 405B Instruct | 405B (dense) | 87.3 MMLU (5-shot) | 50.7 GPQA (0-shot) | 73.8 MATH (CoT) | 89.0 HumanEval | Meta's Llama 3.1 405B Instruct model card |
Each cell names its exact benchmark because the three labs did not run the same test suite: DeepSeek-R1 and Qwen3-235B-A22B report the newer MMLU-Redux / GPQA-Diamond / AIME / LiveCodeBench suite typical of reasoning models, while Llama 3.1 405B (released a year earlier, and not a reasoning model) reports MMLU / GPQA / MATH / HumanEval. Treat scores in different rows of the same column as related, not identical, measurements.
Methodology: how these numbers were compiled
None of the scores above come from a benchmark run Aquanode performed itself. Each one is pulled from the model creator's own technical report, paper or model card, linked in the Source column so you can check the original run yourself. That is a different approach from a live leaderboard that re-runs every model under one harness: it means the number for each model is exactly what its creator claims, at the cost of the three rows not sharing one identical test protocol.
To keep that honest, every cell states the precise benchmark version rather than a generic category name. "MMLU" and "MMLU-Redux" are related but not the same test; the second is a relabeled subset that corrects errors found in the original. The same goes for AIME 2024 versus an older AIME year, or LiveCodeBench versus LiveCodeBench v5. A newer benchmark variant usually exists because the older one had known label errors or had started leaking into training data, so treating the two as interchangeable would overstate how comparable the rows really are.
What each benchmark actually measures: MMLU and MMLU-Redux test broad academic knowledge across dozens of subjects using multiple-choice questions. GPQA and GPQA-Diamond are graduate-level science questions written to resist a quick web search. AIME scores come from actual American Invitational Mathematics Examination problems, a competition math test with a single integer answer per problem. HumanEval and LiveCodeBench both grade a model's generated code by running it against hidden test cases; LiveCodeBench draws its problems from recent programming-contest releases specifically to reduce the chance a model has already seen the answer during training.
Frequently asked questions
What is the best open-source LLM?
It depends on what "best" means for your job. Qwen3-235B-A22B posts the highest scores here on math and competitive coding while activating fewer parameters per token than DeepSeek-R1 (22B vs 37B active), which matters directly for inference cost. DeepSeek-R1 is close behind on most of the same benchmarks and has the widest third-party adoption. Llama 3.1 405B is not a reasoning model in the same sense as the other two (no extended thinking step), so its lower math and coding scores partly reflect that difference in kind, not just raw scale; it remains a strong general-purpose instruct model with the widest ecosystem of fine-tunes and quantizations.
Why do these benchmark scores use different tests per model?
Because that is what each lab actually published. DeepSeek-R1 and Qwen3-235B-A22B report MMLU-Redux, GPQA-Diamond, AIME and LiveCodeBench, the newer benchmark suite reasoning models are evaluated on. Llama 3.1 405B released a year earlier on the older MMLU, GPQA, MATH and HumanEval suite. Reporting the exact benchmark name per cell is more honest than forcing every model onto one column and implying a same-test comparison that never happened.
What do model names like Llama 3.1 405B or Qwen3-235B-A22B mean?
The B suffix is billions of parameters. "405B" is a dense model where all 405 billion parameters run on every token. "235B-A22B" describes a Mixture-of-Experts model: 235 billion total parameters split across experts, but only 22 billion "active" (actually computed) per token, which is why Qwen3-235B-A22B can match or beat a much larger dense model while being cheaper to run. "Instruct" or "Chat" means the base model was further tuned to follow instructions; a model in "thinking" mode has been prompted or tuned to emit an extended reasoning trace before its final answer.
Run any of these models yourself
Use the GPU recommender to size the right card for a model's VRAM footprint, then rent it at the live rate across Aquanode's marketplace of providers.
