From 5816e41cb57e13e52f093b40714e5f57de2e7452 Mon Sep 17 00:00:00 2001 From: sophia Date: Thu, 20 Aug 2026 22:31:45 +0800 Subject: [PATCH 1/2] docs: add Qwen3.8-27B model --- docs/llmservice/models/qwen3.8-27b.md | 43 +++++++++++++++++++ docs/llmservice/pricing-and-usage.md | 1 + .../current/llmservice/models/qwen3.8-27b.md | 43 +++++++++++++++++++ .../current/llmservice/pricing-and-usage.md | 1 + package.json | 2 +- sidebars.js | 1 + 6 files changed, 90 insertions(+), 1 deletion(-) create mode 100644 docs/llmservice/models/qwen3.8-27b.md create mode 100644 i18n/zh-Hans/docusaurus-plugin-content-docs/current/llmservice/models/qwen3.8-27b.md diff --git a/docs/llmservice/models/qwen3.8-27b.md b/docs/llmservice/models/qwen3.8-27b.md new file mode 100644 index 0000000..3679897 --- /dev/null +++ b/docs/llmservice/models/qwen3.8-27b.md @@ -0,0 +1,43 @@ +# Qwen3.8-27B + +## Overview + +Qwen3.8-27B is an open-weight, dense vision-language model from Alibaba's Qwen team with 27 billion parameters. It is designed for coding, professional workflows, research, multimodal interaction, and long-horizon agent tasks. The model accepts text, image, and video input and produces text output. Thinking mode is enabled by default and can be enabled or disabled depending on task requirements. + +## Key Features + +* **Open-Weight Dense Model:** Uses a 27B-parameter dense architecture that balances model capabilities with deployment costs. +* **Ultra-Long Context and Output:** Provides a 262,144-token context window and a 131,072-token maximum output, suitable for large codebases, long documents, and multi-step tasks. +* **Multimodal Input:** Supports text, image, and video inputs with text output. +* **Flexible Reasoning Modes:** Thinking mode is enabled by default, can be disabled when appropriate, and supports different reasoning-effort levels. +* **Agents and Tool Calling:** Supports tool calling, structured output, and multi-step task execution for long-horizon agent workflows. + +## Best Use Cases + +* **Coding and Software Engineering:** Code generation, code comprehension, debugging, refactoring, and repository-level tasks. +* **Professional Workflows:** Complex information organization, structured information extraction, and multi-step business tasks. +* **Research and Long-Document Analysis:** Analysis and summarization of lengthy reports, research papers, knowledge bases, and large-scale contexts. +* **Long-Horizon Agent Tasks:** Tool calling, task planning, sustained reasoning, and multi-turn execution. +* **Multimodal Understanding:** Joint understanding of and question answering across images, videos, and text. + +## Capabilities and Limitations + +| Capability | Description | +| :--- | :--- | +| **Reasoning** | Thinking mode is enabled by default but can be disabled depending on the task; multiple reasoning-effort levels are supported. | +| **Creative Writing** | Suitable for code generation, professional content creation, research, and complex workflows. | +| **Multimodal** | Accepts text, image, and video input and produces text output. | +| **Response Speed** | Response time varies depending on reasoning effort, input length, and output size. | +| **Context Window** | 262,144 tokens. | +| **Max Output** | 131,072 tokens. | +| **Tool Use** | Supports tool calling, structured output, and multi-step task execution. | + +## Pricing + +| Model | Input (Credits/Token) | Cache Write (Credits/Token) | Cache Read (Credits/Token) | Output (Credits/Token) | Web Search (Credits/Use) | +| :--- | --------------------: | --------------------------: | -------------------------: | ---------------------: | -----------------------: | +| **Qwen3.8-27B** | `0.22` | `0.22` | `0.022` | `1.60` | `-` | + +:::info Pricing note +Prices shown in the documentation are B.AI standard reference prices for base billing purposes. B.AI may provide lower actual usage costs through top-up bonuses and account benefits. Specific prices, bonus Credits, and account benefits are subject to the platform display and final billing records. +::: diff --git a/docs/llmservice/pricing-and-usage.md b/docs/llmservice/pricing-and-usage.md index 21e6bd0..594e08c 100644 --- a/docs/llmservice/pricing-and-usage.md +++ b/docs/llmservice/pricing-and-usage.md @@ -25,6 +25,7 @@ The table below lists standard reference prices only. For current limited-time o | Kimi K3 | 3.00 | 3.00 | 0.30 | 15.00 | - | | Kimi K2.6 | 0.95 | 0.95 | 0.1615 | 4.00 | - | | Kimi K2.5 | 0.59 | 0.59 | 0.10 | 3.00 | - | +| Qwen3.8-27B | 0.22 | 0.22 | 0.022 | 1.60 | - | | Qwen3.8-Max | 2.00 | 2.00 | 0.25 | 6.00 | - | | Qwen3.7-Max | 1.65 | 1.65 | 0.33 | 4.951 | - | | Qwen3.6-27B | 0.19 | 0.19 | 0.019 | 2.99 | - | diff --git a/i18n/zh-Hans/docusaurus-plugin-content-docs/current/llmservice/models/qwen3.8-27b.md b/i18n/zh-Hans/docusaurus-plugin-content-docs/current/llmservice/models/qwen3.8-27b.md new file mode 100644 index 0000000..397d00a --- /dev/null +++ b/i18n/zh-Hans/docusaurus-plugin-content-docs/current/llmservice/models/qwen3.8-27b.md @@ -0,0 +1,43 @@ +# Qwen3.8-27B + +## 概述 + +Qwen3.8-27B 是阿里巴巴 Qwen 团队推出的开放权重稠密视觉语言模型，总参数量为 270 亿。该模型面向编程、专业工作流、研究、多模态交互和长周期 Agent 任务；支持文本、图像和视频输入，输出为文本。Thinking 模式默认开启，也可以根据任务需要启用或关闭。 + +## 核心特性 + +* **开放权重稠密模型：** 采用 270 亿参数的稠密架构，在模型能力与部署成本之间取得平衡。 +* **超长上下文与输出：** 支持 262,144 token 上下文窗口和 131,072 token 最大输出，适用于大型代码库、长文档和多步骤任务。 +* **多模态输入：** 支持文本、图像和视频输入，输出为文本。 +* **灵活推理模式：** Thinking 模式默认开启，可按任务需要关闭，并支持不同的推理强度等级。 +* **Agent 与工具调用：** 支持工具调用、结构化输出和多步骤任务执行，适合长周期 Agent 工作流。 + +## 适用场景 + +* **编程与软件工程：** 代码生成、代码理解、调试、重构和仓库级任务。 +* **专业工作流：** 复杂信息整理、结构化信息提取和多步骤业务任务。 +* **研究与长文档分析：** 长篇报告、研究论文、知识库和大规模上下文的分析与总结。 +* **长周期 Agent 任务：** 工具调用、任务规划、持续推理和多轮执行。 +* **多模态理解：** 对图像、视频和文本进行联合理解与问答。 + +## 能力与限制 + +| 能力维度 | 说明 | +| :--- | :--- | +| **推理能力** | Thinking 模式默认开启，但可按任务需要关闭；支持多档推理强度。 | +| **创意与专业内容** | 适合代码生成、专业内容创作、研究和复杂工作流。 | +| **多模态能力** | 支持文本、图像和视频输入，输出为文本。 | +| **响应速度** | 响应时间会受推理强度、输入长度和输出长度影响。 | +| **上下文窗口** | 262,144 token。 | +| **最大输出** | 131,072 token。 | +| **工具调用** | 支持工具调用、结构化输出和多步骤任务执行。 | + +## 价格 + +| 模型名称 | 输入（Credits/Token） | 缓存写入（Credits/Token） | 缓存读取（Credits/Token） | 输出（Credits/Token） | 网页搜索（Credits/次） | +| :--- | --------------------: | -------------------------: | -------------------------: | --------------------: | ---------------------: | +| **Qwen3.8-27B** | `0.22` | `0.22` | `0.022` | `1.60` | `-` | + +:::info 价格说明 +文档价格为 B.AI 平台模型标准参考价，仅供基础计费说明使用。B.AI 可能会通过充值赠送及账户权益等方式，为用户提供更低的实际使用成本。具体价格、赠送积分及账户权益请以平台页面展示及最终账单为准。 +::: diff --git a/i18n/zh-Hans/docusaurus-plugin-content-docs/current/llmservice/pricing-and-usage.md b/i18n/zh-Hans/docusaurus-plugin-content-docs/current/llmservice/pricing-and-usage.md index d0fd9b5..b8f6a06 100644 --- a/i18n/zh-Hans/docusaurus-plugin-content-docs/current/llmservice/pricing-and-usage.md +++ b/i18n/zh-Hans/docusaurus-plugin-content-docs/current/llmservice/pricing-and-usage.md @@ -25,6 +25,7 @@ | Kimi K3 | 3.00 | 3.00 | 0.30 | 15.00 | - | | Kimi K2.6 | 0.95 | 0.95 | 0.1615 | 4.00 | - | | Kimi K2.5 | 0.59 | 0.59 | 0.10 | 3.00 | - | +| Qwen3.8-27B | 0.22 | 0.22 | 0.022 | 1.60 | - | | Qwen3.8-Max | 2.00 | 2.00 | 0.25 | 6.00 | - | | Qwen3.7-Max | 1.65 | 1.65 | 0.33 | 4.951 | - | | Qwen3.6-27B | 0.19 | 0.19 | 0.019 | 2.99 | - | diff --git a/package.json b/package.json index cbf4ba4..6e20581 100644 --- a/package.json +++ b/package.json @@ -1,6 +1,6 @@ { "name": "@x402-tron/docs", - "version": "1.3.23", + "version": "1.3.24", "description": "x402-tron documentation", "license": "MIT", "resolutions": { diff --git a/sidebars.js b/sidebars.js index 0762e89..8c4ebb2 100644 --- a/sidebars.js +++ b/sidebars.js @@ -247,6 +247,7 @@ const sidebars = { label: 'Qwen', collapsed: true, items: [ + { type: 'doc', id: 'llmservice/models/qwen3.8-27b', label: 'Qwen3.8-27B' }, { type: 'doc', id: 'llmservice/models/qwen3.8-max', label: 'Qwen3.8-Max' }, { type: 'doc', id: 'llmservice/models/qwen3.7-max', label: 'Qwen3.7-Max' }, { type: 'doc', id: 'llmservice/models/qwen3.6-27b', label: 'Qwen3.6-27B' }, From 6ee7a371809d814873d9f53dc9ab7f872b2be00c Mon Sep 17 00:00:00 2001 From: sophia Date: Fri, 21 Aug 2026 13:00:31 +0800 Subject: [PATCH 2/2] docs: add MiMo V2.5 models --- docs/llmservice/models/mimo-v2.5-pro.md | 48 +++++++++++++++++++ docs/llmservice/models/mimo-v2.5.md | 48 +++++++++++++++++++ docs/llmservice/pricing-and-usage.md | 2 + .../promotions-and-pricing-notices.md | 2 +- .../current.json | 12 +++++ .../llmservice/models/mimo-v2.5-pro.md | 48 +++++++++++++++++++ .../current/llmservice/models/mimo-v2.5.md | 48 +++++++++++++++++++ .../current/llmservice/pricing-and-usage.md | 2 + .../promotions-and-pricing-notices.md | 2 +- sidebars.js | 9 ++++ 10 files changed, 219 insertions(+), 2 deletions(-) create mode 100644 docs/llmservice/models/mimo-v2.5-pro.md create mode 100644 docs/llmservice/models/mimo-v2.5.md create mode 100644 i18n/zh-Hans/docusaurus-plugin-content-docs/current/llmservice/models/mimo-v2.5-pro.md create mode 100644 i18n/zh-Hans/docusaurus-plugin-content-docs/current/llmservice/models/mimo-v2.5.md diff --git a/docs/llmservice/models/mimo-v2.5-pro.md b/docs/llmservice/models/mimo-v2.5-pro.md new file mode 100644 index 0000000..e94e0e1 --- /dev/null +++ b/docs/llmservice/models/mimo-v2.5-pro.md @@ -0,0 +1,48 @@ +# MiMo-V2.5-Pro + +## Overview + +MiMo-V2.5-Pro is an open-weight text model from Xiaomi MiMo that entered public beta on April 23, 2026. The higher-capacity member of the MiMo-V2.5 family uses a 1.02-trillion-parameter sparse Mixture-of-Experts (MoE) architecture with 42 billion active parameters and is designed for long-horizon agents, software engineering, and complex reasoning. + +## Key Features + +* **1.02T Sparse MoE Architecture:** Uses 1.02 trillion total parameters, 42 billion active parameters, 384 routed experts, and eight selected experts per token. +* **Hybrid Long-Context Attention:** Interleaves 60 Sliding Window Attention layers with 10 global-attention layers and a 128-token sliding window. Xiaomi reports that the 6:1 local-to-global pattern reduces KV-cache storage by nearly 7x. +* **1M Context and 128K Output:** Supports a one-million-token context window and up to 128K output tokens for repository-scale work, long documents, and extended agent trajectories. +* **Agent and Coding Evaluation:** Xiaomi reports 1,581 Elo on GDPVal-AA, 63.8 on ClawEval, 57.2 on SWE-Bench Pro, 78.9 on SWE-Bench Verified, and 68.4 on Terminal-Bench 2.0. +* **Controllable Reasoning and Tools:** Deep thinking is enabled by default but can be disabled. The API supports streaming, function calling, and structured output. + +## Best Use Cases + +* **Long-Horizon Software Engineering:** Repository navigation, multi-file implementation, debugging, testing, and tool-driven work that must remain coherent across long trajectories. +* **Complex Agent Workflows:** Tasks involving repeated function calls, planning, execution, and correction over large working contexts. +* **Long-Document Reasoning:** Analysis and synthesis across extensive code, technical documents, research material, or conversation history. +* **Self-Hosted Model Development:** Commercial inference, fine-tuning, or secondary training where MIT-licensed weights and framework-level deployment control are required. + +## Capabilities and Limitations + +| Capability | Description | +| :--- | :--- | +| **Reasoning** | Deep thinking is enabled by default and can be turned off with `thinking.type`. Xiaomi reports 48.0 on Humanity's Last Exam with tools and 34.0 without tools. | +| **Coding** | Xiaomi reports 57.2 on SWE-Bench Pro, 78.9 on SWE-Bench Verified, 68.4 on Terminal-Bench 2.0, and an implementation rank of 3.4 on FrontierSWE. | +| **Creative Writing** | Supports general and long-form text generation. | +| **Multimodal** | Text input and text output. | +| **Context Window** | 1M tokens. | +| **Max Output** | 128K tokens. | +| **Tool Use** | Supports function calling, structured output, streaming, and Xiaomi's web-search tool. In thinking-mode agent conversations, tool-call history must retain the complete `reasoning_content` field. | +| **Multilingual** | The official model repository identifies English and Chinese support. Xiaomi reports 83.6 on GlobalMMLU, 91.5 on C-Eval, and 90.2 on CMMLU for the Pro Base model, but does not publish a complete language-support list. | + +### Known Limitations + +* `mimo-v2.5-pro` is a text model and does not provide the full-modal understanding available in `mimo-v2.5`. +* In thinking mode, custom `temperature` and `top_p` values are ignored; the API forces `1.0` and `0.95`. Multi-turn tool workflows that omit historical `reasoning_content` can fail with HTTP 400 or lose context quality. + +## Pricing + +| Model | Input (Credits/Token) | Cache Write (Credits/Token) | Cache Read (Credits/Token) | Output (Credits/Token) | Web Search (Credits/Use) | +| :--- | --------------------: | --------------------------: | -------------------------: | ---------------------: | -----------------------: | +| **MiMo-V2.5-Pro** | `0.435` | `0.435` | `0.0036` | `0.87` | `-` | + +:::info Pricing note +Prices shown in the documentation are B.AI standard reference prices for base billing purposes. B.AI may provide lower actual usage costs through top-up bonuses and account benefits. Specific prices, bonus Credits, and account benefits are subject to the platform display and final billing records. +::: diff --git a/docs/llmservice/models/mimo-v2.5.md b/docs/llmservice/models/mimo-v2.5.md new file mode 100644 index 0000000..878e039 --- /dev/null +++ b/docs/llmservice/models/mimo-v2.5.md @@ -0,0 +1,48 @@ +# MiMo-V2.5 + +## Overview + +MiMo-V2.5 is an open-weight native omnimodal model from Xiaomi MiMo that entered public beta on April 23, 2026. It combines a 310-billion-parameter sparse Mixture-of-Experts (MoE) language backbone with dedicated vision and audio encoders to understand text, images, video, and audio in long-context agent workflows. + +## Key Features + +* **Native Omnimodal Understanding:** Processes text, image, video, and audio input in one model, using a 729M-parameter vision encoder and a 261M-parameter audio encoder in addition to the language backbone. +* **310B Sparse MoE Architecture:** Uses 310 billion total parameters, 15 billion active parameters, 256 routed experts, and eight selected experts per token. +* **Efficient 1M-Token Context:** Interleaves 39 Sliding Window Attention layers with nine global-attention layers in a 5:1 pattern. Xiaomi reports nearly 6x lower KV-cache storage while retaining long-context performance. +* **Multimodal and Coding Evaluation:** Xiaomi reports 77.9 on MMMU-Pro, 87.7 on Video-MME, 23.8 on Claw-Eval Multimodal, 65.8 on Terminal-Bench 2.0, and 56.1 on SWE-Bench Pro. +* **Agent-Oriented API Controls:** Supports deep thinking, streaming, function calling, structured output, and Xiaomi's separately billed web-search service. + +## Best Use Cases + +* **Multimodal Analysis:** Understanding screenshots, charts, scanned documents, video, and audio alongside written instructions and contextual data. +* **Visual and Media-Aware Agents:** Agents that must inspect multimodal evidence, reason about it, call tools, and produce a text response or action plan. +* **General Coding and Automation:** Software tasks that benefit from image-based UI context, terminal tools, structured output, and lower API cost than the Pro variant. +* **Long-Context Applications:** Document analysis, codebase assistance, media archives, and extended conversations that need up to a one-million-token context. + +## Capabilities and Limitations + +| Capability | Description | +| :--- | :--- | +| **Reasoning** | Deep thinking is enabled by default and can be turned off with `thinking.type`. Xiaomi's official multimodal results include 77.9 on MMMU-Pro and 81.0 on CharXiv RQ. | +| **Coding** | Xiaomi reports 71.8 on MiMo Coding Bench, 62.3 on Claw-Eval Text, 65.8 on Terminal-Bench 2.0, and 56.1 on SWE-Bench Pro. | +| **Creative Writing** | Supports general and long-form text generation. | +| **Multimodal** | Accepts text, image, video, and audio input and produces text output. Official results include 87.7 on Video-MME, 83.5 on DailyOmni, and 23.8 on Claw-Eval Multimodal. | +| **Context Window** | 1M tokens. | +| **Max Output** | 128K tokens. | +| **Tool Use** | Supports function calling, structured output, streaming, and Xiaomi's web-search tool. In thinking-mode agent conversations, tool-call history must retain the complete `reasoning_content` field. | +| **Multilingual** | The official model repository identifies English and Chinese support. | + +### Known Limitations + +* The hosted API provides multimodal understanding with text output; it does not generate images, audio, or video as native response modalities. +* In thinking mode, custom `temperature` and `top_p` values are ignored; the API forces `1.0` and `0.95`. Multi-turn tool workflows that omit historical `reasoning_content` can fail with HTTP 400 or lose context quality. + +## Pricing + +| Model | Input (Credits/Token) | Cache Write (Credits/Token) | Cache Read (Credits/Token) | Output (Credits/Token) | Web Search (Credits/Use) | +| :--- | --------------------: | --------------------------: | -------------------------: | ---------------------: | -----------------------: | +| **MiMo-V2.5** | `0.14` | `0.14` | `0.0028` | `0.28` | `-` | + +:::info Pricing note +Prices shown in the documentation are B.AI standard reference prices for base billing purposes. B.AI may provide lower actual usage costs through top-up bonuses and account benefits. Specific prices, bonus Credits, and account benefits are subject to the platform display and final billing records. +::: diff --git a/docs/llmservice/pricing-and-usage.md b/docs/llmservice/pricing-and-usage.md index 594e08c..0324b26 100644 --- a/docs/llmservice/pricing-and-usage.md +++ b/docs/llmservice/pricing-and-usage.md @@ -30,6 +30,8 @@ The table below lists standard reference prices only. For current limited-time o | Qwen3.7-Max | 1.65 | 1.65 | 0.33 | 4.951 | - | | Qwen3.6-27B | 0.19 | 0.19 | 0.019 | 2.99 | - | | Hy3 | 0.132 | 0.132 | 0.033 | 0.528 | - | +| MiMo-V2.5-Pro | 0.435 | 0.435 | 0.0036 | 0.87 | - | +| MiMo-V2.5 | 0.14 | 0.14 | 0.0028 | 0.28 | - | | GLM-5.3 | 1.40 | 1.40 | 0.28 | 4.40 | - | | GLM-5.2 | 1.40 | 1.40 | 0.28 | 4.40 | - | | GLM-5.1 | 1.40 | 1.40 | 0.28 | 4.40 | - | diff --git a/docs/llmservice/promotions-and-pricing-notices.md b/docs/llmservice/promotions-and-pricing-notices.md index 5c85689..3f30f41 100644 --- a/docs/llmservice/promotions-and-pricing-notices.md +++ b/docs/llmservice/promotions-and-pricing-notices.md @@ -32,7 +32,7 @@ After the offer ends, DeepSeek-V4-Flash will return to standard pricing. See the status="Free Offer" detail="0 Credits" > -Offer starts August 20, 2026. +Offer starts August 21, 2026. This offer applies to B.AI Chat and API: diff --git a/i18n/zh-Hans/docusaurus-plugin-content-docs/current.json b/i18n/zh-Hans/docusaurus-plugin-content-docs/current.json index d5e427f..009f336 100644 --- a/i18n/zh-Hans/docusaurus-plugin-content-docs/current.json +++ b/i18n/zh-Hans/docusaurus-plugin-content-docs/current.json @@ -91,6 +91,18 @@ "message": "Hy3", "description": "The label for the Hy3 model document in sidebar docsSidebar" }, + "sidebar.docsSidebar.category.Xiaomi (MiMo)": { + "message": "小米（MiMo）", + "description": "The label for the Xiaomi MiMo model category in sidebar docsSidebar" + }, + "sidebar.docsSidebar.doc.MiMo-V2.5-Pro": { + "message": "MiMo-V2.5-Pro", + "description": "The label for the MiMo-V2.5-Pro model document in sidebar docsSidebar" + }, + "sidebar.docsSidebar.doc.MiMo-V2.5": { + "message": "MiMo-V2.5", + "description": "The label for the MiMo-V2.5 model document in sidebar docsSidebar" + }, "sidebar.docsSidebar.category.OpenClaw": { "message": "OpenClaw 集成", "description": "The label for category OpenClaw in sidebar docsSidebar" diff --git a/i18n/zh-Hans/docusaurus-plugin-content-docs/current/llmservice/models/mimo-v2.5-pro.md b/i18n/zh-Hans/docusaurus-plugin-content-docs/current/llmservice/models/mimo-v2.5-pro.md new file mode 100644 index 0000000..5f4926a --- /dev/null +++ b/i18n/zh-Hans/docusaurus-plugin-content-docs/current/llmservice/models/mimo-v2.5-pro.md @@ -0,0 +1,48 @@ +# MiMo-V2.5-Pro + +## 概述 + +MiMo-V2.5-Pro 是小米 MiMo 于 2026 年 4 月 23 日进入公开测试的开放权重文本模型。作为 MiMo-V2.5 系列中容量更高的成员，它采用 1.02 万亿参数的稀疏 Mixture-of-Experts（MoE）架构，激活参数量为 420 亿，面向长周期 Agent、软件工程和复杂推理任务。 + +## 核心特性 + +* **1.02T 稀疏 MoE 架构：** 总参数量为 1.02 万亿，激活参数量为 420 亿，包含 384 个路由专家，每个 token 选择其中 8 个专家。 +* **混合长上下文注意力：** 交错使用 60 层滑动窗口注意力和 10 层全局注意力，滑动窗口为 128 token。小米表示，6:1 的局部与全局层比例可将 KV Cache 存储降低近 7 倍。 +* **1M 上下文与 128K 输出：** 支持 100 万 token 上下文窗口和最高 128K token 输出，适合仓库级工作、长文档和长周期 Agent 轨迹。 +* **Agent 与编程评测：** 小米报告其在 GDPVal-AA、ClawEval、SWE-Bench Pro、SWE-Bench Verified 和 Terminal-Bench 2.0 上的结果分别为 1,581 Elo、63.8、57.2、78.9 和 68.4。 +* **可控推理与工具：** 深度思考默认开启，但可以关闭；API 支持流式输出、函数调用和结构化输出。 + +## 适用场景 + +* **长周期软件工程：** 在长轨迹中保持连贯地进行仓库导航、多文件实现、调试、测试和工具驱动工作。 +* **复杂 Agent 工作流：** 在大规模工作上下文中重复执行函数调用、规划、执行和修正的任务。 +* **长文档推理：** 跨大量代码、技术文档、研究资料或对话历史进行分析与综合。 +* **自托管模型开发：** 适用于需要基于 MIT 许可权重进行商业推理、微调或二次训练，并需获得框架级部署控制的场景。 + +## 能力与限制 + +| 能力维度 | 说明 | +| :--- | :--- | +| **推理能力** | 深度思考默认开启，可通过 `thinking.type` 关闭。小米报告其在 Humanity's Last Exam 上使用工具得分 48.0、未使用工具得分 34.0。 | +| **编程能力** | 小米报告其在 SWE-Bench Pro、SWE-Bench Verified、Terminal-Bench 2.0 上的得分分别为 57.2、78.9、68.4，并在 FrontierSWE 实现排名中取得 3.4。 | +| **创意写作** | 支持通用和长篇文本生成。 | +| **多模态能力** | 仅支持文本输入和文本输出。 | +| **上下文窗口** | 1M token。 | +| **最大输出** | 128K token。 | +| **工具调用** | 支持函数调用、结构化输出、流式输出和小米网页搜索工具。思考模式的 Agent 对话中，工具调用历史需保留完整的 `reasoning_content` 字段。 | +| **多语言能力** | 官方模型仓库标注支持英文和中文。小米报告 Pro Base 模型在 GlobalMMLU、C-Eval 和 CMMLU 上分别为 83.6、91.5、90.2，但未公布完整语言覆盖范围。 | + +### 已知限制 + +* `mimo-v2.5-pro` 是文本模型，不提供 `mimo-v2.5` 具备的全模态理解能力。 +* 在思考模式下，自定义 `temperature` 和 `top_p` 不生效，API 固定使用 `1.0` 和 `0.95`。多轮工具工作流若遗漏历史 `reasoning_content`，可能返回 HTTP 400 或导致上下文质量下降。 + +## 价格 + +| 模型名称 | 输入（Credits/Token） | 缓存写入（Credits/Token） | 缓存读取（Credits/Token） | 输出（Credits/Token） | 网页搜索（Credits/次） | +| :--- | --------------------: | -------------------------: | -------------------------: | --------------------: | ---------------------: | +| **MiMo-V2.5-Pro** | `0.435` | `0.435` | `0.0036` | `0.87` | `-` | + +:::info 价格说明 +文档价格为 B.AI 平台模型标准参考价，仅供基础计费说明使用。B.AI 可能会通过充值赠送及账户权益等方式，为用户提供更低的实际使用成本。具体价格、赠送积分及账户权益请以平台页面展示及最终账单为准。 +::: diff --git a/i18n/zh-Hans/docusaurus-plugin-content-docs/current/llmservice/models/mimo-v2.5.md b/i18n/zh-Hans/docusaurus-plugin-content-docs/current/llmservice/models/mimo-v2.5.md new file mode 100644 index 0000000..8f33682 --- /dev/null +++ b/i18n/zh-Hans/docusaurus-plugin-content-docs/current/llmservice/models/mimo-v2.5.md @@ -0,0 +1,48 @@ +# MiMo-V2.5 + +## 概述 + +MiMo-V2.5 是小米 MiMo 于 2026 年 4 月 23 日进入公开测试的开放权重原生全模态模型。它将一个 3100 亿参数的稀疏 Mixture-of-Experts（MoE）语言骨干网络与专用视觉、音频编码器结合，可在长上下文 Agent 工作流中理解文本、图像、视频和音频。 + +## 核心特性 + +* **原生全模态理解：** 在同一模型中处理文本、图像、视频和音频输入；除语言骨干网络外，还包含 7.29 亿参数视觉编码器和 2.61 亿参数音频编码器。 +* **310B 稀疏 MoE 架构：** 总参数量为 3100 亿，激活参数量为 150 亿，包含 256 个路由专家，每个 token 选择其中 8 个专家。 +* **高效 1M Token 上下文：** 以 5:1 的模式交错使用 39 层滑动窗口注意力和 9 层全局注意力。小米表示，在保留长上下文能力的同时，KV Cache 存储可降低近 6 倍。 +* **多模态与编程评测：** 小米报告其在 MMMU-Pro、Video-MME、Claw-Eval Multimodal、Terminal-Bench 2.0 和 SWE-Bench Pro 上的得分分别为 77.9、87.7、23.8、65.8 和 56.1。 +* **面向 Agent 的 API 控制：** 支持深度思考、流式输出、函数调用、结构化输出，以及小米单独计费的网页搜索服务。 + +## 适用场景 + +* **多模态分析：** 结合书面指令和上下文数据，理解截图、图表、扫描文档、视频和音频。 +* **视觉与媒体 Agent：** 检查多模态信息、对其推理、调用工具，并生成文本回复或行动计划。 +* **通用编程与自动化：** 利用图像化 UI 上下文、终端工具和结构化输出完成软件任务，相比 Pro 版本具有更低的 API 成本。 +* **长上下文应用：** 支持文档分析、代码库辅助、媒体档案和最长 100 万 token 的长对话。 + +## 能力与限制 + +| 能力维度 | 说明 | +| :--- | :--- | +| **推理能力** | 深度思考默认开启，可通过 `thinking.type` 关闭。小米官方多模态结果包括 MMMU-Pro 77.9 和 CharXiv RQ 81.0。 | +| **编程能力** | 小米报告其在 MiMo Coding Bench、Claw-Eval Text、Terminal-Bench 2.0 和 SWE-Bench Pro 上的得分分别为 71.8、62.3、65.8 和 56.1。 | +| **创意写作** | 支持通用和长篇文本生成。 | +| **多模态能力** | 支持文本、图像、视频和音频输入，输出为文本。官方结果包括 Video-MME 87.7、DailyOmni 83.5 和 Claw-Eval Multimodal 23.8。 | +| **上下文窗口** | 1M token。 | +| **最大输出** | 128K token。 | +| **工具调用** | 支持函数调用、结构化输出、流式输出和小米网页搜索工具。思考模式的 Agent 对话中，工具调用历史需保留完整的 `reasoning_content` 字段。 | +| **多语言能力** | 官方模型仓库标注支持英文和中文。 | + +### 已知限制 + +* 托管 API 提供多模态理解和文本输出，不原生生成图像、音频或视频。 +* 在思考模式下，自定义 `temperature` 和 `top_p` 不生效，API 固定使用 `1.0` 和 `0.95`。多轮工具工作流若遗漏历史 `reasoning_content`，可能返回 HTTP 400 或导致上下文质量下降。 + +## 价格 + +| 模型名称 | 输入（Credits/Token） | 缓存写入（Credits/Token） | 缓存读取（Credits/Token） | 输出（Credits/Token） | 网页搜索（Credits/次） | +| :--- | --------------------: | -------------------------: | -------------------------: | --------------------: | ---------------------: | +| **MiMo-V2.5** | `0.14` | `0.14` | `0.0028` | `0.28` | `-` | + +:::info 价格说明 +文档价格为 B.AI 平台模型标准参考价，仅供基础计费说明使用。B.AI 可能会通过充值赠送及账户权益等方式，为用户提供更低的实际使用成本。具体价格、赠送积分及账户权益请以平台页面展示及最终账单为准。 +::: diff --git a/i18n/zh-Hans/docusaurus-plugin-content-docs/current/llmservice/pricing-and-usage.md b/i18n/zh-Hans/docusaurus-plugin-content-docs/current/llmservice/pricing-and-usage.md index b8f6a06..c6bfd45 100644 --- a/i18n/zh-Hans/docusaurus-plugin-content-docs/current/llmservice/pricing-and-usage.md +++ b/i18n/zh-Hans/docusaurus-plugin-content-docs/current/llmservice/pricing-and-usage.md @@ -30,6 +30,8 @@ | Qwen3.7-Max | 1.65 | 1.65 | 0.33 | 4.951 | - | | Qwen3.6-27B | 0.19 | 0.19 | 0.019 | 2.99 | - | | Hy3 | 0.132 | 0.132 | 0.033 | 0.528 | - | +| MiMo-V2.5-Pro | 0.435 | 0.435 | 0.0036 | 0.87 | - | +| MiMo-V2.5 | 0.14 | 0.14 | 0.0028 | 0.28 | - | | GLM-5.3 | 1.40 | 1.40 | 0.28 | 4.40 | - | | GLM-5.2 | 1.40 | 1.40 | 0.28 | 4.40 | - | | GLM-5.1 | 1.40 | 1.40 | 0.28 | 4.40 | - | diff --git a/i18n/zh-Hans/docusaurus-plugin-content-docs/current/llmservice/promotions-and-pricing-notices.md b/i18n/zh-Hans/docusaurus-plugin-content-docs/current/llmservice/promotions-and-pricing-notices.md index 0e1ad63..8e0fae9 100644 --- a/i18n/zh-Hans/docusaurus-plugin-content-docs/current/llmservice/promotions-and-pricing-notices.md +++ b/i18n/zh-Hans/docusaurus-plugin-content-docs/current/llmservice/promotions-and-pricing-notices.md @@ -32,7 +32,7 @@ import ActivityCard from '@site/src/components/ActivityCard'; status="免费活动" detail="0 Credits" > -活动开始时间：2026 年 8 月 20 日。 +活动开始时间：2026 年 8 月 21 日。 本活动覆盖 B.AI Chat 和 API： diff --git a/sidebars.js b/sidebars.js index 8c4ebb2..ed733e2 100644 --- a/sidebars.js +++ b/sidebars.js @@ -213,6 +213,15 @@ const sidebars = { { type: 'doc', id: 'llmservice/models/hy3', label: 'Hy3' }, ], }, + { + type: 'category', + label: 'Xiaomi (MiMo)', + collapsed: true, + items: [ + { type: 'doc', id: 'llmservice/models/mimo-v2.5-pro', label: 'MiMo-V2.5-Pro' }, + { type: 'doc', id: 'llmservice/models/mimo-v2.5', label: 'MiMo-V2.5' }, + ], + }, { type: 'category', label: 'SpaceXAI (Grok)',