Home › Benchmarks
Kimi K3.1 Pro Benchmarks
Kimi K3.1 Pro entered the Artificial Analysis composite leaderboard at #4 when Moonshot AI released it on 23 October 2026. It is a 3T-parameter open-weight model with a 1M-token context window, and the launch notes report a 65% reduction in hallucination rate versus Kimi K3. Below are the scores, what each benchmark measures, and how to read them without being misled.
① 「Benchmark scores」表的逐项分数（官方 model card 或 Artificial Analysis 各分项页面）
② 「Artificial Analysis positioning」里的综合分值与前后位次
③ 「Long-context performance」的实测衰减数据
④ 若官方未公布某项，直接删那一行——不要留空行，更不要估
All figures on this page are dated. Last synced: 24 October 2026. Primary source: Moonshot AI official release notes for Kimi K3.1 Pro, published 23 October 2026; positioning cross-checked against Artificial Analysis on 24 October 2026. Benchmark results are not directly comparable across evaluators — see how to read these numbers before drawing conclusions.
Kimi K3.1 Pro benchmark scores
The table below lists each benchmark separately rather than collapsing everything into one number. Composite ranks are convenient but they hide the fact that two models can score identically overall while being far apart on the task you actually run in production.
| Benchmark | What it measures | K3.1 Pro | K3 | Delta |
|---|---|---|---|---|
| MMLU-Pro | Broad academic knowledge across 14 domains, 10-option format | — | — | — |
| GPQA Diamond | Graduate-level physics, chemistry and biology; resists lookup | — | — | — |
| AIME 2025 | Competition mathematics, answer-only scoring | — | — | — |
| SWE-bench Verified | Resolving real GitHub issues in Python repositories | — | — | — |
| LiveCodeBench | Contest programming problems published after training cutoff | — | — | — |
| τ²-bench | Tool use and multi-turn agentic task completion | — | — | — |
Scores are reported as percentages. Delta is K3.1 Pro minus K3 in percentage points. Where a benchmark is not reported for one of the two models, the row is omitted rather than filled with an estimate.
Artificial Analysis positioning
Artificial Analysis publishes one of the few independent, continuously refreshed comparisons that covers both closed and open-weight models on the same scale. On its composite leaderboard, Kimi K3.1 Pro entered at #4 at launch — the single most useful external signal, because the vendor did not produce it.
artificialanalysis.ai 搜 Kimi K3.1 Pro，抄当期数字并记下抓取日期——这个榜每月都在动，
写死「#4」但不说日期，两周后就会变成错信息。
How to read Kimi K3.1 Pro benchmarks without being misled
1. 同一分数在不同评测方之间不可比
MMLU-Pro 在 A 家跑出来的 78 和在 B 家跑出来的 78 不是同一件事——提示词模板、few-shot 数量、是否允许思维链、答案抽取规则，任何一项变了分数就会动好几个点。**同一张表里只跟同一来源的数字比。** 本页每张表都标了来源和同步日期，就是为了让你能判断可比性。
2. 竞赛数学和真实工程能力之间隔着一道墙
AIME 类分数高，说明模型在「答案唯一的短推理题」上强；SWE-bench 高，说明它能在真实代码库里定位问题、改对、跑过测试。这两件事的相关性远低于多数人以为的水平。做代码工具的，看 SWE-bench 和 LiveCodeBench；做 Agent 的，看 τ²-bench；做知识问答的，才看 MMLU-Pro。
3. 上下文窗口长度 ≠ 长上下文能力
1M token 是允许塞多少，不是塞满后还能答多准。几乎所有模型在长上下文里都存在「中间遗忘」，越靠中间的信息越容易被跳过。判断一个模型的长上下文能力，要看「在 100K / 500K / 1M 各档位的准确率曲线」，而不是看标称窗口大小。
4. 基准污染是所有公开榜单的通病
SWE-bench、MMLU 这类老基准的题目早已大面积进入训练语料，「背下来的分数」和「真会做的分数」无法从公开数字里区分。这解释了为什么新发布的模型在新基准（比如 cutoff 之后才出的 LiveCodeBench 题）上表现往往明显低于老基准。看到一个高得离谱的老基准分数时，先怀疑污染。
What actually changed from Kimi K3
K3.1 Pro 是 K3 的改进版，不是新架构。发布会强调的提升集中在三处，都很务实：
| 维度 | K3.1 Pro | 说明 |
|---|---|---|
| API 价格 | $2 / $10 per 1M | 比 K3 的 $3 / $15 便宜约 33%，输入输出两侧同降（详见 定价页） |
| 幻觉率 | −65% | 官方口径相对 K3 的下降幅度，是可核实的少数硬指标之一 |
| 上下文窗口 | 1,000,000 | 与 K3 持平，本代没有扩展 |
| 参数规模 | 3T | 开源权重，可自部署 |
| Highspeed 变体 | 有 | 低延迟服务版本，与标准版分开计价 |
关于网络安全能力：发布说明提到本代在该类能力上做了主动克制，因此相关评测分数会低于「若无限制」的预期值。这是有意的产品选择，不是能力倒退——引用这类分数时务必说明背景，否则会得出完全相反的结论。
Where to verify these numbers yourself
Frequently asked questions
What are Kimi K3.1 Pro's benchmark scores?
At launch Kimi K3.1 Pro ranked #4 on the Artificial Analysis composite leaderboard.
Individual benchmark scores differ by test, so the tables above list each benchmark separately alongside the source it was taken from, rather than collapsing everything into one headline number. Where the official model card reports a score for both K3.1 Pro and K3, the delta column shows the change in percentage points.
How does Kimi K3.1 Pro compare with Kimi K3 on benchmarks?
Kimi K3.1 Pro is a refinement of K3 rather than a new architecture. The launch notes present the gains as better reliability and long-context behaviour — including a reported 65% reduction in hallucination rate — rather than a large jump in raw capability. In practice that means you should expect modest movement on most academic benchmarks and clearer improvement on tasks where a wrong-but-confident answer is the failure mode.
What is the context window of Kimi K3.1 Pro?
1,000,000 tokens, unchanged from Kimi K3. Long-context benchmarks are reported separately from short-context ones because accuracy typically degrades as the input grows, so a single context-length figure does not tell you how well the model holds up across an entire window. If your workload routinely fills most of the window, test it on your own documents before committing.
Is Kimi K3.1 Pro open weights, and can I self-host it?
Kimi K3.1 Pro is described as a 3T-parameter open-weight model, so the weights are published. Serving a model of that size requires a multi-GPU deployment and is not something you run on a single workstation. Check the licence terms in the official release notes before any commercial deployment — "open weights" and "permissive licence" are not the same thing.
Is Kimi K3.1 Pro better than GPT or Claude?
That depends on the benchmark and on which version you are comparing. Kimi K3.1 Pro entered the Artificial Analysis composite at #4 at launch, which places it among the strongest models available, but a composite rank hides wide variation between tasks. Compare the per-benchmark rows for the task you actually care about, and where possible run your own evaluation set — public benchmarks tell you the shape of a model's strengths, not whether it will work on your data.
How often are the benchmark numbers on this page updated?
Every table carries a sync date. Scores move when the vendor publishes a new model card, when an independent evaluator re-runs its suite, and when third-party leaderboards refresh. Figures here are re-checked against the primary source before each update, and where a change is material the previous value is noted rather than silently overwritten.
Related on this site
Pricing is the other half of the picture — a benchmark lead that costs three times as much per token is not a lead in production. The Kimi K3.1 Pro pricing page breaks down the official $2 / $10 per-million-token rate against third-party providers and includes a token cost calculator, so you can price a workload at the same time as you size up the model.
Spotted an error, or a score that has moved since our last sync? This page is maintained by hand and corrections are welcome — accuracy matters more to us than looking right.