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. | 671B | 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. | 128k | 8k |
| Benchmarks | ||
LiveCodeBenchCompetitive coding — Coding problems published so recently the AI can't have seen them in training — a contamination-free test of raw programming skill. Higher is better. | 88.7% | 14.3% |
GPQA DiamondScience — Graduate-level science questions in biology, physics, and chemistry — hard enough that subject-matter PhDs score around 65%. Higher is better. | 85.7% | 35.2% |
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. | — | 48.1% |
Humanity's Last Exam · text only, no toolsMultidisciplinary reasoning — accuracy on the 2,158 text-only questions without tools | 30.6% | — |
MATHReasoning — accuracy on the original competition mathematics test set | — | 45.3% |
WinoGrandeReasoning — resolving ambiguous pronouns in context | — | 74.6% |
AIME 2025Reasoning — single-answer accuracy on the 2025 invitational mathematics exam without tools | 96% | — |
SimpleQAKnowledge — accuracy on short, fact-seeking questions in the original SimpleQA set | — | 5.6% |
MMLU-ProKnowledge — accuracy on the ten-option MMLU-Pro benchmark | — | 42.4% |
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. | — | 70% |
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. | — | 77.4% |
| Overview | ||
| Company | DeepSeek | Moonshot AI |
| Release date | Dec 1 2025 | Feb 22 2025 |
| Access | Open Weight | Open Weight |
| Model details | View model | View model |
Other comparisons
DeepSeek-V3.2-SpecialevsClaude Fable 5.2MoonlightvsClaude Fable 5.2DeepSeek-V3.2-SpecialevsGPT-6.2 AstraMoonlightvsGPT-6.2 AstraDeepSeek-V3.2-SpecialevsGemini 4 HydrogenMoonlightvsGemini 4 HydrogenDeepSeek-V3.2-SpecialevsMuse EmberMoonlightvsMuse EmberDeepSeek-V3.2-SpecialevsGrok 4.7MoonlightvsGrok 4.7DeepSeek-V3.2-SpecialevsMistral Large 4MoonlightvsMistral Large 4Frequently asked questions
DeepSeek-V3.2-Speciale leads Moonlight on 2 of the 2 benchmarks they both report (LiveCodeBench, GPQA Diamond). Moonlight shipped 282 days before DeepSeek-V3.2-Speciale, so benchmark comparisons should account for the intervening progress.
DeepSeek-V3.2-Speciale has 671B parameters, while Moonlight has 16B. Context windows are 128k (DeepSeek-V3.2-Speciale) vs 8k (Moonlight).
On LiveCodeBench, DeepSeek-V3.2-Speciale leads at 88.7% vs Moonlight at 14.3%. On GPQA Diamond, DeepSeek-V3.2-Speciale leads at 85.7% vs Moonlight at 35.2%.