MoonlightOpen Weight

Released

Moonlight is an AI model released by Moonshot AI on Saturday, Feb 22 2025, 33 days after Kimi K1.5. It is an open-weight model — the trained weights are available to download and run. It is a 16B parameterThe 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. model with a 8k token 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.. Benchmark results (shown below) cover LiveCodeBench, HumanEval, MATH, WinoGrande, GPQA Diamond, SimpleQA, and 3 more.

Benchmarks

Coding

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.
14.3%
#70 of 70
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%
#62 of 74

Reasoning & science

MATHReasoning — accuracy on the original competition mathematics test set
45.3%
#35 of 50
WinoGrandeReasoning — resolving ambiguous pronouns in context
74.6%
#8 of 9
GPQA DiamondScience — Graduate-level science questions in biology, physics, and chemistry — hard enough that subject-matter PhDs score around 65%. Higher is better.
35.2%
#128 of 129
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%
#54 of 68
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%
#25 of 39

Knowledge work

SimpleQAKnowledge — accuracy on short, fact-seeking questions in the original SimpleQA set
5.6%
#21 of 21
MMLU-ProKnowledge — accuracy on the ten-option MMLU-Pro benchmark
42.4%
#37 of 37

Compare Moonlight with

Moonlight

About

Moonshot AI introduced Moonlight in February 2025 as a compact mixture-of-experts language model for studying the Muon optimizer at a larger training scale. The team trained it on 5.7 trillion tokens and released pretrained, instruction-tuned and intermediate checkpoints so others could examine how it learned.

The pretrained checkpoint scored 70.0% on MMLU and 48.1% on HumanEval in Moonshot's release report. Moonshot later evaluated the same base checkpoint on GPQA Diamond and LiveCodeBench v6 for its Kimi Linear report. An instruction-tuned version shipped alongside the base model for conversational use.

Frequently asked questions

Moonlight was released by Moonshot AI on Saturday, Feb 22 2025.

All Moonshot AI releases

15 tracked

2023

1 release