Kimi's open model K3 nears GPT-5.6 Sol and Fable 5 while signaling the end of super cheap Chinese AI

Kimi 正在推出 K3，这是一款参数量为 2.8 万亿、具有一百万令牌上下文窗口的多模态模型。在公司自己的基准测试中，它的表现与领先的专有模型相当。

据 Kimi 介绍，新旗舰模型 K3 共有 2.8 万亿参数，能够原生处理图像和视频，并支持一百万令牌的上下文窗口。Kimi 称 K3 是大约 3 万亿参数范围内的首个开放模型。完整的模型权重预计于 7 月 27 日发布。该模型主要面向长期运行的编程任务、知识工作和复杂推理。

在 Kimi 自己的基准测试中：https://www.kimi.com/blog/kimi-k3，K3 仍然落后于顶级专有模型 Claude Fable 5 和 GPT 5.6 Sol，但击败了所有其他测试系统，包括 Claude Opus 系列模型和中国竞争对手 GLM-5.2。根据公司介绍，所有结果均来自 Kimi，并在最大或高强度思考下获得。

在所有 35 个测试中，K3 大约有七次排名第一，在其余大部分测试中排名第二或第三。Fable 5 赢得了最多的单项测试。在几乎每个基准测试中，K3 都以较大优势击败了 Opus 4.8、GPT 5.5 和 GLM 5.2。根据不同的基准测试，使用了三种代理系统中的一种：KimiCode、Claude Code 或 Codex。这意味着结果并非全部在相同条件下收集。

独立测试实验室 Artificial Analysis：https://x.com/ArtificialAnlys/status/2077832874183860404 已发布其对 Kimi K3 的首次评估。该模型在 Artificial Analysis 智能指数中的得分为 57，与 Opus 4.8 和 GPT-5.5 相当，但仍落后于 Fable 5 和 GPT-5.6 Sol。这与 Kimi 自己的说法大致一致。

在代理任务中，K3 在 GDPval v2 上的 Elo 评分达到 1,668，相较 K2.6 的 1,190 是一个大幅提升。它击败了 GLM-5.2（1,514）、GPT-5.5（1,494）和 Claude Opus 4.8（1,600），但仍未超过 Claude Fable 5（1,760）。在 AutomationBench-AA——Artificial Analysis 对 Zapier 代理 SaaS 工作流程评估的版本——中，K3 同样位居榜首，得分为 53%。

在 AA-Briefcase（一个私人长期知识工作评估）中，K3 达到了 1,547 的整体 Elo 分，比 K2.6 高出 732 分。只有 Claude Fable 5 得分更高。Artificial Analysis 评价 K3 表现全面，其评分标准和分析质量接近 Fable 5 的水平。不过，GPT-5.6 Sol 在展示质量上仍然领先。

K3 在 AA-Omniscience 指数上的准确率从 33% 提高到 46%，将整体分数从 +6 提升至 +18。但其幻觉率也从 39% 上升到 51%，意味着即使 K3 回答更多问题正确，它也会编造更多答案。Ad DEC_D_Incontent-2

根据 Kimi 的说法，该模型的主要使用场景是长期软件开发，且几乎无需人工监督。K3 旨在分析大型代码库、协调终端工具，并在多个工作步骤中保持对任务的专注。Ad

该模型将编程与视觉反馈相结合：它会检查屏幕截图、修改代码，然后检查可见输出。Kimi 将这种闭环系统称为“Vision in the Loop”，并将其定位为游戏开发、UI 设计和 CAD 的基础。

作为演示，Kimi 展示了一个程序生成的 3D 开放世界游戏：https://horseback-open-world.ok.kimi.link/，据称 K3 完全在浏览器中使用 Three.js、WebGPU 和 GPU Compute 构建，以及一个交互式黑洞可视化：https://blackhole-visualizer.ok.kimi.link/。对于开放世界演示，K3 程序生成了环境，并使用外部工具创建了 3D 骑手和马模型。其他演示：https://www.kimi.com/blog/kimi-k3 包括长征十号火箭发射与返回的模拟，以及一个 Game Boy Advance 模拟器。

K3 使用专家混合架构，每次仅激活 896 个专家中的 16 个。它配备了一个名为 Kimi Delta Attention 的新注意力架构：https://arxiv.org/abs/2510.26692，Kimi 称该架构可实现百万 token 上下文的解码速度提高至 6.3 倍。所谓的“注意力残差”：https://arxiv.org/abs/2603.15031 据称可将训练效率提高约 25%，同时额外计算开销不足 2%。

根据 Kimi API 文档：https://platform.kimi.com/，一百万输入令牌的成本为缓存命中时 $0.30，未命中时 $3.00。包括推理在内的一百万输出令牌成本为 $15.00。这些价格与上下文长度无关。缓存会自动进行，这使得未修改的长前缀对于代理和大型代码库特别有用。

这使得 K3 的价格远高于其前代 K2.6。K2.6 官方价格为缓存命中时每百万令牌 $0.16，未命中时 $0.95，输出为 $4.00。中国供应商也不再以极低价格提供其前沿模型。

不过，K3 仍比最顶级的西方模型便宜，处于上中档价格区间。例如，Anthropic 的新模型 Sonnet 5：https://the-decoder.com/claude-sonnet-5-continues-anthropics-pattern-of-hiding-price-increases-behind-unchanged-token-rates/，输入令牌每百万 $3，输出 $15，但性能较低。

根据 Artificial Analysis：https://x.com/ArtificialAnlys/status/2077832885021835289，K3 在智能指数上的平均每任务成本为 $0.94，接近 GPT-5.6 Sol 的 $1.04，大约是 Opus 4.8 $1.80 的一半。然而，它明显高于开源权重的同类模型，如 GLM-5.2（$0.32）和 DeepSeek V4 Pro（$0.04）。

K3 的令牌使用量也比其前代少。完成所有九项评估大约需要 1.32 亿输出令牌，而 K2.6 约需 1.66 亿，减少了 21%，同时得分高出 13 分。由于令牌价格大幅提升，在大多数情况下，K3 每任务的成本仍可能高于 K2.6。

K3 已可通过 Kimi.com：http://kimi.com/，iOS、Android 和 HarmonyOS 移动应用，Kimi Work 桌面客户端：https://www.kimi.com/zh-cn/products/kimi-work（版本 3.1.0 及以上），以及 Kimi Code：https://www.kimi.com/code 获取。在 OpenRouter 上，该模型的标识为 "moonshotai/kimi-k3"，但目前只通过 Moonshot 本身提供服务。预计开源权重将在七月底前发布。

保持对 AI 的了解，内容清晰、有用，无冗余。

关注 The Decoder 获取 AI 新闻、背景故事和专家分析。

The Decoder：https://the-decoder.com/

Kimi is launching K3, a multimodal model with 2.8 trillion parameters and a context window of one million tokens. In the company's own benchmarks, it performs on par with leading proprietary models.

According to Kimi, the new flagship model K3 has 2.8 trillion total parameters, processes images and video natively, and supports a context window of one million tokens. Kimi calls K3 the first open model in the roughly 3 trillion parameter range. Full model weights are scheduled for release by July 27. The model targets long-running programming tasks, knowledge work, and complex reasoning.

In Kimi's own benchmarks：https://www.kimi.com/blog/kimi-k3, K3 still trails the top proprietary models Claude Fable 5 and GPT 5.6 Sol but beats every other system tested, including the Claude Opus models and Chinese rival GLM-5.2. All results come from Kimi and were achieved at maximum or high thinking intensity, according to the company. Ad

Across all 35 tests, K3 took first place about seven times and landed second or third in most of the rest. Fable 5 won the most individual tests. In nearly every benchmark, K3 beat Opus 4.8, GPT 5.5, and GLM 5.2 by a wide margin. Depending on the benchmark, one of three agent systems was used: KimiCode, Claude Code, or Codex. That means the results weren't all collected under identical conditions. Ad DEC_D_Incontent-1

Independent testing lab Artificial Analysis：https://x.com/ArtificialAnlys/status/2077832874183860404 has published its first evaluation of Kimi K3. The model scores 57 on the Artificial Analysis Intelligence Index, putting it on par with Opus 4.8 and GPT-5.5 but still behind Fable 5 and GPT-5.6 Sol. That largely lines up with Kimi's own claims.

On agentic tasks, K3 reaches an Elo rating of 1,668 on GDPval v2, a big jump from K2.6's 1,190. It beats GLM-5.2 (1,514), GPT-5.5 (1,494), and Claude Opus 4.8 (1,600), though it still falls short of Claude Fable 5 (1,760). K3 also takes the top spot on AutomationBench-AA, Artificial Analysis's version of Zapier's agentic SaaS workflow evaluation, with a score of 53 percent. Ad

On AA-Briefcase, a private long-horizon knowledge work evaluation, K3 reaches an overall Elo of 1,547, up 732 points from K2.6. Only Claude Fable 5 scores higher. Artificial Analysis calls K3 well-rounded, with rubric scoring and analytical quality close to Fable 5's level. GPT-5.6 Sol still leads on presentation quality, though.

K3's accuracy rate improved from 33 percent to 46 percent on the AA-Omniscience Index, pushing the overall score from +6 to +18. But its hallucination rate climbed from 39 percent to 51 percent, meaning K3 fabricates more answers even as it gets more questions right. Ad DEC_D_Incontent-2

According to Kimi, the model's primary use case is long-running software development with minimal human oversight. K3 is built to analyze large codebases, coordinate terminal tools, and stay focused on a task across many work steps. Ad

The model pairs programming with visual feedback: it examines screen captures, modifies code, then checks the visible output. Kimi calls this closed-loop system "Vision in the Loop" and positions it as a foundation for game development, UI design, and CAD.

As demos, Kimi shows off a procedurally generated 3D open-world game：https://horseback-open-world.ok.kimi.link/ that K3 reportedly built entirely in the browser using Three.js, WebGPU, and GPU Compute, along with an interactive black hole visualization：https://blackhole-visualizer.ok.kimi.link/. For the open-world demo, K3 procedurally generated the environment and used an external tool to create the 3D rider and horse models. Other demos：https://www.kimi.com/blog/kimi-k3 include a simulation of the Long March 10 rocket launch and return, plus a Game Boy Advance emulator.

K3 uses a mixture-of-experts architecture that activates only 16 of 896 experts at a time. It's paired with a new attention architecture called Kimi Delta Attention：https://arxiv.org/abs/2510.26692, which Kimi says enables up to 6.3x faster decoding for million-token contexts. "Attention residuals"：https://arxiv.org/abs/2603.15031 reportedly boost training efficiency by about 25 percent while adding less than 2 percent in extra compute overhead.

According to the Kimi API docs：https://platform.kimi.com/, one million input tokens cost $0.30 with a cache hit and $3.00 without. One million output tokens, including reasoning, cost $15.00. These prices apply regardless of context length. Caching happens automatically, which makes unmodified long prefixes especially useful for agents and large codebases.

That puts K3 well above the price level of its predecessor K2.6, which officially costs $0.16 per million tokens with a cache hit, $0.95 without, and $4.00 for output. Chinese providers aren't offering their frontier models at rock-bottom prices anymore either.

Still, K3 is much cheaper than the top Western models and sits more in the upper midrange. Anthropic's new Sonnet 5：https://the-decoder.com/claude-sonnet-5-continues-anthropics-pattern-of-hiding-price-increases-behind-unchanged-token-rates/, for example, also costs $3 per million input tokens and $15 for output but delivers lower performance.

According to Artificial Analysis：https://x.com/ArtificialAnlys/status/2077832885021835289, K3 averages $0.94 per task on the Intelligence Index, close to GPT-5.6 Sol at $1.04 and about half the price of Opus 4.8 at $1.80. It's well above open-weight peers like GLM-5.2 ($0.32) and DeepSeek V4 Pro ($0.04), though.

K3 also uses fewer tokens than its predecessor. It needed about 132 million output tokens to complete all nine evaluations, down from roughly 166 million for K2.6, a 21 percent reduction while scoring 13 points higher. Because of the much higher token prices, K3 will likely still cost more per task than K2.6 in most cases.

K3 is already available through Kimi.com：http://kimi.com/, the mobile app for iOS, Android, and HarmonyOS, the Kimi Work desktop：https://www.kimi.com/zh-cn/products/kimi-work client (version 3.1.0 and later), and Kimi Code：https://www.kimi.com/code. On OpenRouter, the model is listed under the identifier "moonshotai/kimi-k3," though it's currently served there only through Moonshot itself. The open weights are expected by the end of July.

Stay in the loop on AI. Clear, useful, no fluff.

Follow The Decoder for AI news, background stories and expert analyses.

The Decoder：https://the-decoder.com/