Kimi K3、DeepSeek V4 Pro 与 GLM-5.2 三款开源万亿参数 MoE 模型对比

目前三家中国实验室占据了开源重量级排行榜的前列。Moonshot AI 的 Kimi K3：https://www.kimi.com/blog/kimi-k3、DeepSeek V4 Pro：https://api-docs.deepseek.com/news/news260424/ 和 Zhipu AI 的 GLM-5.2：https://huggingface.co/zai-org/GLM-5.2 都是稀疏专家混合（MoE）模型，具有百万令牌的上下文窗口。每个模型都面向长期编码和代理工作负载。本文从 AI 团队实际决策的三个维度对它们进行比较：测量能力、许可证条款和服务成本。

“Kimi K3（2.8T）和 DeepSeek V4 Pro（1.6T）”都属于“万亿参数”级别。GLM-5.2 总参数为 7440 亿，因此在三者中总参数最小。它之所以能够占据一席之地，是因为在 K3 发布之前，它已经引领了开源领域。

Kimi K3：https://www.kimi.com/blog/kimi-k3 是一个 2.8 万亿参数的稳定 LatentMoE 模型，每个令牌激活 896 个专家中的 16 个。Moonshot 尚未公布确切的活跃参数数量。K3 增加了原生视觉功能、100 万令牌的上下文窗口，以及持续开启的推理功能。Moonshot 称其为首个开放的 3 万亿级模型。我们的发布报道在此：https://www.marktechpost.com/2026/07/16/moonshot-ai-releases-kimi-k3-a-2-8-trillion-parameter-open-moe-model-with-kimi-delta-attention-and-1m-context/。

DeepSeek V4 Pro：https://api-docs.deepseek.com/news/news260424/ 是一个拥有 1.6 万亿参数、49B 活跃参数的 MoE 模型，使用 384 个路由专家和 1 个共享专家。它具有 100 万令牌的上下文窗口和 38.4 万的最大输出。较小的 V4 Flash 变体（总参数 284B，活跃参数 13B）适用于更廉价的工作负载。权重可在 Hugging Face 获取：https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro。

GLM-5.2：https://huggingface.co/zai-org/GLM-5.2 是一个 7440 亿参数的 MoE 模型，活跃参数约为 400 亿，拥有 100 万令牌的上下文窗口。智谱科技提供了 High 和 Max 两种推理模式，并附带 API 访问。

厂商报告的评分使用了不同的测试框架，因此各基准数值在不同实验室之间很少完全一致。中立的比较参考是人工分析智能指数：https://artificialanalysis.ai/models/comparisons/kimi-k3-vs-deepseek-v4-pro，该指数在同一套评测中对三者进行评分。

在该指数上，Kimi K3 得分约为 57：https://artificialanalysis.ai/articles/kimi-k3-achieves-3-in-the-artificial-analysis-intelligence-index-comparable-to-opus-4-8-and-gpt-5-5，DeepSeek V4 Pro（最大推理能力）得分为 44，GLM-5.2 得分为 51。K3 排名总体第三，仅次于 Claude Fable 5 和 GPT-5.6 Sol，并与 Opus 4.8 和 GPT-5.5 相当。GLM-5.2 在 K3 发布前保持开放权重的首位。

编码基准测试显示了类似的情况，但要注意一些注意事项。Moonshot 自己的表格通过匹配的测试环境运行 K3 和 GLM-5.2。在那里，K3 在所有共享基准上均以较大幅度领先 GLM-5.2。

DeepSeek 在 Moonshot 的表格中未出现，因此其数据来自单独测试。DeepSeek-V4-Pro-Max 在 SWE-bench Verified 上得分 80.6%：https://www.morphllm.com/deepseek-v4，为发布时最高的开放权重结果，并与 Gemini 3.1 Pro 并列。它在 MRCR 1M 上得分 83.5，确认其强大的长上下文能力。GLM-5.2 在 SWE-bench Pro 上得分 62.1：https://kie.ai/blog/glm-5-2-benchmark-deep-dive，略高于 GPT-5.5 的 58.6 分。

因此，在测量能力上，K3 是三者中最强的。DeepSeek V4 Pro 在单独的编码任务上具有竞争力。GLM-5.2 落后于 K3，但仍然是一个有能力的开放权重选择。

三者都作为开放权重模型发布，但目前实际状况有所不同。

DeepSeek V4 Pro 是 MIT 许可：https://huggingface.co/collections/deepseek-ai/deepseek-v4，从第一天起权重就在 Hugging Face 上可用。GLM-5.2 也是 MIT 许可：https://huggingface.co/zai-org/GLM-5.2，在 zai-org 组织下 Hugging Face 上提供完整权重。两者都允许无限制的商业使用、微调和自托管。

Kimi K3 是例外。Moonshot 承诺将在 2026 年 7 月 27 日前发布权重，预计采用修改版 MIT 许可。在此之前，K3 只能通过 API 和 Kimi 应用使用。Moonshot 最近的修改版 MIT 条款增加了一个归属条款，仅在每月活跃用户超过 1 亿时触发。

API 定价将这些模型明显区分开来。

DeepSeek V4 Pro 是成本领先者，优势明显。在列表输出率下，1 美元大约可以从 V4 Pro 购买 1.15M 输出令牌，从 GLM-5.2 大约购买 227K，从 K3 大约购买 67K。

Artificial Analysis 按照 7:2:1 的缓存/输入/输出加权基准对每个模型进行定价，这样可以消除供应商的框架影响。在此基础上，它列出了 K3 每百万 tokens 价格为 $2.31，GLM-5.2 为 $0.90，而 DeepSeek V4 Pro 为 $0.18。按每个任务的成本，同一来源报告 K3 为 $0.94，GLM-5.2 为 $0.32，DeepSeek V4 Pro 为 $0.04：https://artificialanalysis.ai/articles/kimi-k3-achieves-3-in-the-artificial-analysis-intelligence-index-comparable-to-opus-4-8-and-gpt-5-5。

速度也有所不同。Artificial Analysis 测得 GLM-5.2 约为每秒 168 tokens：https://artificialanalysis.ai/models/comparisons/glm-5-2-vs-deepseek-v4-pro，远超过 DeepSeek V4 Pro 和 Kimi K3 两者的约每秒 62 tokens。Moonshot 报告在编码工作负载中缓存命中率超过 90%，这将 K3 的有效输入成本降至每百万 tokens $0.30。

自托管是另一个限制。GLM-5.2 744B 在 BF16 下需要超过 1TB 的显存，或大约 8 块 FP8 H200 显卡。DeepSeek V4 Pro 1.6T 需求更高。Kimi K3 是最重的：Moonshot 建议使用 64 块或更多加速器，这使得大多数团队无法本地部署。K3 使用 MXFP4 权重和 MXFP8 激活以支持更广泛的硬件。

Three Chinese labs now hold the top of the open-weight leaderboard. Moonshot AI’s Kimi K3：https://www.kimi.com/blog/kimi-k3, DeepSeek V4 Pro：https://api-docs.deepseek.com/news/news260424/, and Zhipu AI’s GLM-5.2：https://huggingface.co/zai-org/GLM-5.2 are all sparse Mixture-of-Experts (MoE) models with million-token context windows. Each targets long-horizon coding and agent workloads. This article compares them on three axes an AI team actually decides on: measured capability, license terms, and serving cost.

‘Trillion-parameter’ fits Kimi K3 (2.8T) and DeepSeek V4 Pro (1.6T). GLM-5.2 is 744B total, so it is the smallest of the three by total parameters. It earns its place because it led the open-weight field before K3 shipped.

Kimi K3：https://www.kimi.com/blog/kimi-k3 is a 2.8-trillion-parameter Stable LatentMoE model activating 16 of 896 experts per token. Moonshot has not published the exact active-parameter count. K3 adds native vision, a 1M-token context window, and always-on reasoning. Moonshot calls it the first open 3T-class model. Our launch coverage is here：https://www.marktechpost.com/2026/07/16/moonshot-ai-releases-kimi-k3-a-2-8-trillion-parameter-open-moe-model-with-kimi-delta-attention-and-1m-context/.

DeepSeek V4 Pro：https://api-docs.deepseek.com/news/news260424/ is a 1.6-trillion-parameter MoE with 49B active parameters, using 384 routed experts plus one shared expert. It carries a 1M-token context window with 384K max output. A smaller V4 Flash variant (284B total, 13B active) covers cheaper workloads. Weights are on Hugging Face：https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro.

GLM-5.2：https://huggingface.co/zai-org/GLM-5.2 is a 744-billion-parameter MoE with roughly 40B active parameters and a 1M-token context window. Zhipu ships it with High and Max reasoning modes. It comes with API access

Vendor-reported scores use different harnesses, so per-benchmark numbers rarely line up cleanly across labs. The neutral comparator is the Artificial Analysis Intelligence Index：https://artificialanalysis.ai/models/comparisons/kimi-k3-vs-deepseek-v4-pro, which scores all three on the same suite.

On that index, Kimi K3 scores about 57：https://artificialanalysis.ai/articles/kimi-k3-achieves-3-in-the-artificial-analysis-intelligence-index-comparable-to-opus-4-8-and-gpt-5-5, DeepSeek V4 Pro (Max reasoning) scores 44, and GLM-5.2 scores 51. K3 ranks #3 overall, behind only Claude Fable 5 and GPT-5.6 Sol, and comparable to Opus 4.8 and GPT-5.5. GLM-5.2 held the top open-weight spot until K3 shipped.

Coding benchmarks tell a similar story with caveats. Moonshot’s own table runs K3 and GLM-5.2 through matched harnesses. There, K3 leads GLM-5.2 on every shared benchmark by wide margins.

DeepSeek does not appear in Moonshot’s table, so its numbers come from separate testing. DeepSeek-V4-Pro-Max scores 80.6% on SWE-bench Verified：https://www.morphllm.com/deepseek-v4, the highest open-weight result at its release and tied with Gemini 3.1 Pro. It also posts 83.5 on MRCR 1M, confirming serious long-context ability. GLM-5.2 scored 62.1 on SWE-bench Pro：https://kie.ai/blog/glm-5-2-benchmark-deep-dive, edging GPT-5.5 at 58.6.

So, K3 is the strongest of the three on measured capability. DeepSeek V4 Pro is competitive on isolated coding tasks. GLM-5.2 trails K3 but remains a capable open-weight option.

All three ship as open-weight models, but the practical status differs today.

DeepSeek V4 Pro is MIT-licensed：https://huggingface.co/collections/deepseek-ai/deepseek-v4, with weights on Hugging Face from day one. GLM-5.2 is also MIT-licensed：https://huggingface.co/zai-org/GLM-5.2, with full weights on Hugging Face under the zai-org organization. Both allow unrestricted commercial use, fine-tuning, and self-hosting now.

Kimi K3 is the exception. Moonshot has committed to publishing weights by July 27, 2026, expected under a Modified MIT license. Until then, K3 is usable only through the API and Kimi apps. Moonshot’s recent Modified MIT terms add one attribution clause. It triggers only above 100M monthly active users.

API list pricing separates these models sharply.

DeepSeek V4 Pro is the cost leader by a wide margin. At list output rates, one dollar buys roughly 1.15M output tokens from V4 Pro, about 227K from GLM-5.2, and about 67K from K3.

Artificial Analysis prices every model on one blended 7:2:1 cache/input/output basis, which removes vendor framing. On that basis it lists K3 at $2.31 per 1M tokens, GLM-5.2 at $0.90, and DeepSeek V4 Pro at $0.18. On cost per task, the same source reports K3 at $0.94, GLM-5.2 at $0.32, and DeepSeek V4 Pro at $0.04：https://artificialanalysis.ai/articles/kimi-k3-achieves-3-in-the-artificial-analysis-intelligence-index-comparable-to-opus-4-8-and-gpt-5-5.

Speed also differs. Artificial Analysis measures GLM-5.2 at about 168 tokens/sec：https://artificialanalysis.ai/models/comparisons/glm-5-2-vs-deepseek-v4-pro, well ahead of DeepSeek V4 Pro and Kimi K3 at about 62 each. Moonshot reports above 90% cache hits in coding workloads, which drops K3’s effective input cost to $0.30 per million.

Self-hosting is a different constraint. GLM-5.2 at 744B needs over 1TB of VRAM in BF16, or roughly 8x H200 at FP8. DeepSeek V4 Pro at 1.6T needs more still. Kimi K3 is heaviest: Moonshot recommends 64 or more accelerators, putting local serving out of reach for most teams. K3 uses MXFP4 weights with MXFP8 activations for broader hardware support.