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. | 16B | — |
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. | 32k | 1M |
| 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. | 57.3% | 74.3% |
MATHReasoning — accuracy on the original competition mathematics test set | 27.9% | 54.9% |
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. | 55.7% | 78.9% |
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. | 72% | 86.2% |
SWE-Bench VerifiedCoding — Real coding tasks pulled from open-source projects — the AI has to find and fix actual bugs. A human-checked version of the original SWE-Bench. Higher is better. | — | 9.6% |
HellaSwagReasoning — commonsense completion of short scenarios | — | 86.5% |
MGSMMultilingual math — solving grade-school word problems translated into multiple languages | — | 82.6% |
MathVistaMultimodal reasoning — mathematical reasoning about visual information | — | 58.4% |
MMMUMultimodal — Tests the AI on understanding images and text together across many college subjects. Higher is better. | — | 56.1% |
| Overview | ||
| Company | DeepSeek | |
| Release date | May 16 2024 | May 14 2024 |
| Access | Open Weight | Closed |
| Model details | View model | View model |
Other comparisons
DeepSeek V2 LitevsClaude Fable 5.2Gemini 1.5 FlashvsClaude Fable 5.2DeepSeek V2 LitevsGPT-6.2 AstraGemini 1.5 FlashvsGPT-6.2 AstraDeepSeek V2 LitevsMuse EmberGemini 1.5 FlashvsMuse EmberDeepSeek V2 LitevsGrok 4.7Gemini 1.5 FlashvsGrok 4.7DeepSeek V2 LitevsMistral Large 4Gemini 1.5 FlashvsMistral Large 4DeepSeek V2 LitevsKimi K3.1 ProGemini 1.5 FlashvsKimi K3.1 ProFrequently asked questions
Gemini 1.5 Flash leads DeepSeek V2 Lite on 4 of the 4 benchmarks they both report (HumanEval, MATH, MMLU, GSM8K). Gemini 1.5 Flash shipped 2 days before DeepSeek V2 Lite, so benchmark comparisons should account for the intervening progress.
Context windows are 32k (DeepSeek V2 Lite) vs 1M (Gemini 1.5 Flash). DeepSeek V2 Lite is open weight, while Gemini 1.5 Flash is closed.
On HumanEval, Gemini 1.5 Flash leads at 74.3% vs DeepSeek V2 Lite at 57.3%. On MATH, Gemini 1.5 Flash leads at 54.9% vs DeepSeek V2 Lite at 27.9%. On MMLU, Gemini 1.5 Flash leads at 78.9% vs DeepSeek V2 Lite at 55.7%. On GSM8K, Gemini 1.5 Flash leads at 86.2% vs DeepSeek V2 Lite at 72%.