[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"$f1jbami8rt7qkh":3},{"slug":4,"category":5,"publishedAt":6,"titleZh":7,"titleEn":8,"summaryZh":9,"summaryEn":10,"models":11,"hardwares":24,"bodyZh":25,"bodyEn":26,"sourceName":27,"sourceUrl":28,"updatedAt":29},"kimi-k31-pro-2026-10-24","NEWS","2026-10-24T10:00:00.000Z","Kimi K3.1 Pro 发布：2.8T-A104B、1M 上下文、每百万 token 输入 $2","Kimi K3.1 Pro ships: 2.8T-A104B, a 1M-token context and $2 per million input tokens","10 月 23 日月之暗面发布 Kimi K3 系列首个模型 K3.1 Pro——K3 的后训练版，2.8T 总参 \u002F 104B 激活，1,048,576 token 上下文，原生多模态。Terminal-Bench 4.0 拿到 65.7%（高于 GPT-6 Astra 的 58.2%）。官方定价输入 $2.00 \u002F 输出 $10.00（每百万 token），较前代 K3 降约 33%；另发 2-3 倍速的 Highspeed 版（$4 \u002F $20）。权重承诺 11 月 4 日前放出，站内按 RESTRICTED 口径收录，零实测记录。","Moonshot AI released Kimi K3.1 Pro on Oct 23, the K3 series' first model — a post-trained K3 with 2.8T total \u002F 104B active parameters, a 1M-token context and native multimodality. It scores 65.7% on Terminal-Bench 4.0, ahead of GPT-6 Astra at 58.2%. Pricing is $2.00 input \u002F $10.00 output per million tokens, about 33% below the previous K3, plus a 2-3x faster Highspeed build at $4 \u002F $20. Weights are promised by Nov 4, so the site logs it under the RESTRICTED convention with zero runs.",[12,15,18,21],{"id":13,"name":14},47,"Qwen3.6-35B-A3B",{"id":16,"name":17},73,"Kimi-K3",{"id":19,"name":20},98,"Kimi-K3.1-Pro",{"id":22,"name":23},99,"Kimi-K3.1-Pro-Highspeed",[],"# Kimi K3.1 Pro 发布：2.8T-A104B、1M 上下文、每百万 token 输入 $2\n\n10 月 23 日，月之暗面（Moonshot AI）发布 **Kimi K3.1 Pro**——K3 系列的第一个模型，也是 K3 的**后训练版本**：官方称「迄今能力最强的 3T 级开放权重模型」，参数规格沿用 K3 的 **2.8T 总参 \u002F 104B 激活**，上下文 **1,048,576 token（1M）**，原生多模态（文本 \u002F 图像 \u002F 视频）。社区热度也印证了它的分量：HN 上这条 713 分、453 条评论，是本窗口最热的 LLM 话题。\n\n## 基准：Terminal-Bench 4.0 65.7%\n\n| 基准 | K3.1 Pro | 对照 |\n|---|---|---|\n| Terminal-Bench 4.0 | **65.7%** | GPT-6 Astra 58.2% \u002F Claude Opus 5.5 66.4% |\n| Kimi Code Bench 3.0 Long-horizon | 61.2% | 前代 K3 55.95% |\n| BenchCAD（带工具） | 92.7% | Claude Opus 5.5 96.2% \u002F GPT-6 Astra 95.9% |\n| OfficeQA Pro v2 | **64.4** | 官方称 SOTA |\n\n后训练是这版的重点：官方描述为「跨数百个下游领域的大规模联合 RL」，为此重写了 RL 框架、长上下文推理、训练效率、沙箱、rollout 调度与环境质量检查整条栈。3D 建模与金融能力同步升级——BenchCAD 上能写出 Blender \u002F CAD 脚本，OfficeQA Pro v2 是文档级财务推理。\n\n## 定价：比前代降约 33%\n\n官方 API 定价（每百万 token）：\n\n| 模型 | 输入 | 输出 | 缓存输入 | 缓存写入 5m \u002F 1h |\n|---|---|---|---|---|\n| **kimi-k3.1-pro** | **$2.00** | **$10.00** | $0.20 | $2.00 \u002F $4.00 |\n| **kimi-k3.1-pro-highspeed** | $4.00 | $20.00 | $0.20 | $4.00 \u002F $8.00 |\n| kimi-k3（前代） | $3.00 | $15.00 | $0.30 | $3.00 \u002F $6.00 |\n\nHighspeed 版是同一套权重的加速服务档（官方称 **2-3 倍**速度），价格正好翻倍。API 今日可用（模型名 `kimi-k3.1-pro`），Kimi.ai \u002F Kimi Work \u002F Kimi Code 同步上线；推理档位由 `reasoning_effort` 控制（low \u002F high \u002F max，默认 max）。\n\n## 权重 11 月 4 日前放出\n\n**当前只有 API，权重未放出**——官方承诺「full model weights will be released by Nov 4」。站内按既有口径收录：\n\n| 模型 | ID | 状态 | 实测 |\n|---|---|---|---|\n| Kimi-K3.1-Pro | 98 | RESTRICTED | 零记录 |\n| Kimi-K3.1-Pro-Highspeed | 99 | RESTRICTED | 零记录 |\n| Kimi-K3（前代，2.8T-A104B） | 73 | 参数已补全 | 56 条 |\n\n顺带把 K3 与 Gemma-4-E4B-it 的参数字段补全（此前为空）：K3 = 2.8T 总参 \u002F 104B 激活，gemma-4-E4B-it = 8B 总参 \u002F 4.5B 有效参。\n\n**本地部署的现实**：2.8T 权重即使按 NVFP4 \u002F MXFP4 也是 TB 级，跟上一批次 Flint（39.8B，GGUF Q4_K_XL 25.1GB）、Raon-1-mini（28.8B，Q4_K_M 19.7GB）不是一个量级。站内给 2.8T 模型判 fit 的结果一贯是「不 fit」——K3 那 56 条实测也全部来自更早的轻量口径，不构成对 K3.1 的速度背书。11 月 4 日放权重之后，MXFP4 权重 + 高速投机解码的组合值不值得折腾，得看官方仓库的量化件厚度。","# Kimi K3.1 Pro ships: 2.8T-A104B, a 1M-token context and $2 per million input tokens\n\nOn Oct 23, Moonshot AI released **Kimi K3.1 Pro**, the first model in the K3 series and a **post-trained version of K3**: billed as the company's most capable 3T-class open-weight model to date, it keeps K3's **2.8T total \u002F 104B active** parameters, a **1,048,576-token (1M) context window** and native multimodality (text\u002Fimage\u002Fvideo). Community attention matched: the thread took 713 points and 453 comments on HN, the hottest LLM story of the window.\n\n## Benchmarks: 65.7% on Terminal-Bench 4.0\n\n| Benchmark | K3.1 Pro | Comparison |\n|---|---|---|\n| Terminal-Bench 4.0 | **65.7%** | GPT-6 Astra 58.2% \u002F Claude Opus 5.5 66.4% |\n| Kimi Code Bench 3.0 Long-horizon | 61.2% | previous K3 55.95% |\n| BenchCAD (with tools) | 92.7% | Claude Opus 5.5 96.2% \u002F GPT-6 Astra 95.9% |\n| OfficeQA Pro v2 | **64.4** | claimed state of the art |\n\nPost-training is the point of this release: Moonshot describes it as large-scale joint RL across hundreds of downstream domains, which meant rewriting the RL framework, long-context inference, training efficiency, sandbox infrastructure, rollout scheduling and a shared environment-quality platform. 3D creation and finance improved in step — on BenchCAD it writes Blender and CAD scripts, and OfficeQA Pro v2 tests document-level fiscal reasoning.\n\n## Pricing: about 33% below the previous generation\n\nOfficial API pricing per million tokens:\n\n| Model | Input | Output | Cached input | Cache write 5m \u002F 1h |\n|---|---|---|---|---|\n| **kimi-k3.1-pro** | **$2.00** | **$10.00** | $0.20 | $2.00 \u002F $4.00 |\n| **kimi-k3.1-pro-highspeed** | $4.00 | $20.00 | $0.20 | $4.00 \u002F $8.00 |\n| kimi-k3 (previous) | $3.00 | $15.00 | $0.30 | $3.00 \u002F $6.00 |\n\nThe Highspeed build is the same weights served faster (officially 2-3x) at exactly double the price. The API is live today as `kimi-k3.1-pro`, alongside Kimi.ai, Kimi Work and Kimi Code; the reasoning tier is chosen with `reasoning_effort` (low \u002F high \u002F max, max by default).\n\n## Weights by Nov 4\n\n**API only for now — the weights are not out**, and the official line is that full model weights will be released by Nov 4. The site logs it under the established convention:\n\n| Model | ID | Status | Runs |\n|---|---|---|---|\n| Kimi-K3.1-Pro | 98 | RESTRICTED | zero |\n| Kimi-K3.1-Pro-Highspeed | 99 | RESTRICTED | zero |\n| Kimi-K3 (previous, 2.8T-A104B) | 73 | params backfilled | 56 |\n\nWe also backfilled two previously blank parameter fields: K3 (2.8T total \u002F 104B active) and gemma-4-E4B-it (8B total \u002F 4.5B effective).\n\n**The local-deployment reality**: 2.8T parameters is terabytes even at NVFP4\u002FMXFP4, a different order of magnitude from last batch's Flint (39.8B, GGUF Q4_K_XL 25.1GB) or Raon-1-mini (28.8B, Q4_K_M 19.7GB). The site's fit verdict for a 2.8T model is consistently \"does not fit\" — and the 56 existing K3 rows predate this release, so they say nothing about K3.1's speed. Whether the Nov 4 weight drop makes MXFP4 plus fast speculative decoding worth the trouble depends entirely on how thick the official quantized artifacts are.","araoai","https:\u002F\u002Faraoai.com\u002Fnews\u002Fkimi-k31-pro-2026-10-24","2026-10-24T01:46:21.421Z"]