Compare AI models
| Specifications | ||
ParametersThe learned values that shape how a model responds. B means billion and T means trillion. For mixture-of-experts models, the total count is larger than the number active for each token. Parameter count alone does not measure quality. | 236B | — |
Context windowHow much information the model can work with in one request, measured in tokens — pieces of words or other data. This includes your question, attachments, conversation and reply. Limits can vary by model variant and provider. | 128k | 32k |
| Benchmarks | ||
HumanEvalFunction synthesis — 164 small Python problems: the AI is given a function's description and has to write the working function. This was the coding benchmark of the GPT-3.5 and GPT-4 era, before the field moved to fixing real bugs in real repositories. Higher is better. | 84.76% | 74.4% |
MATHReasoning — accuracy on the original competition mathematics test set | 71.02% | 53.2% |
HellaSwagReasoning — commonsense completion of short scenarios | — | 87.8% |
MMLUGeneral knowledge — A 57-subject multiple-choice exam — history, law, medicine, maths — that was the standard measure of how much a model knows from 2020 until roughly 2024, when frontier scores crowded into the high 80s and labs moved on to harder tests. The scores here were published years apart under different testing setups, so read them as a historical record rather than a like-for-like ranking. Higher is better. | — | 83.7% |
MGSMMultilingual math — solving grade-school word problems translated into multiple languages | — | 79% |
GSM8KGrade-school math — Grade-school maths word problems that take a few steps of arithmetic to work through. It separated the models of 2022 and 2023 sharply, then saturated. One caveat on the historical numbers: OpenAI included part of the GSM8K training set in GPT-4's pre-training mix, so GPT-4's score is not a clean few-shot result. Higher is better. | — | 88.9% |
MathVistaMultimodal reasoning — mathematical reasoning about visual information | — | 53% |
MMMUMultimodal — Tests the AI on understanding images and text together across many college subjects. Higher is better. | — | 59.4% |
| Overview | ||
| Company | DeepSeek | |
| Release date | Jun 28 2024 | Feb 8 2024 |
| Access | Open Weight | Closed |
| Model details | View model | View model |
Other comparisons
DeepSeek V2-0628vsClaude Fable 5.2Gemini 1.0 UltravsClaude Fable 5.2DeepSeek V2-0628vsGPT-6.2 AstraGemini 1.0 UltravsGPT-6.2 AstraDeepSeek V2-0628vsMuse EmberGemini 1.0 UltravsMuse EmberDeepSeek V2-0628vsGrok 4.7Gemini 1.0 UltravsGrok 4.7DeepSeek V2-0628vsMistral Large 4Gemini 1.0 UltravsMistral Large 4DeepSeek V2-0628vsKimi K3.1 ProGemini 1.0 UltravsKimi K3.1 ProFrequently asked questions
DeepSeek V2-0628 leads Gemini 1.0 Ultra on 2 of the 2 benchmarks they both report (HumanEval, MATH). Gemini 1.0 Ultra shipped 141 days before DeepSeek V2-0628, so benchmark comparisons should account for the intervening progress.
Context windows are 128k (DeepSeek V2-0628) vs 32k (Gemini 1.0 Ultra). DeepSeek V2-0628 is open weight, while Gemini 1.0 Ultra is closed.
On HumanEval, DeepSeek V2-0628 leads at 84.76% vs Gemini 1.0 Ultra at 74.4%. On MATH, DeepSeek V2-0628 leads at 71.02% vs Gemini 1.0 Ultra at 53.2%.