diff --git a/mintlify-docs/ai-tools/build-local-rgb-agent.mdx b/mintlify-docs/ai-tools/build-local-rgb-agent.mdx
index 619f017..baca55c 100644
--- a/mintlify-docs/ai-tools/build-local-rgb-agent.mdx
+++ b/mintlify-docs/ai-tools/build-local-rgb-agent.mdx
@@ -108,11 +108,11 @@ npm init -y && npm pkg set type=module
npm install @kaleidorg/mind @qvac/sdk @modelcontextprotocol/sdk
```
-KaleidoMind works with `@qvac/sdk` 0.13 and later; installing the latest release is recommended. Save this as `agent.mjs`:
+KaleidoMind works with `@qvac/sdk` 0.13.1 and later and is tested with 0.21. Save this as `agent.mjs`:
```js agent.mjs
import { createInterface } from 'node:readline/promises';
-import { completion, cancel, loadModel, QWEN3_600M_INST_Q4 } from '@qvac/sdk';
+import { completion, cancel, loadModel, QWEN3_5_2B_MULTIMODAL_Q4_K_M } from '@qvac/sdk';
import { Funnel, ToolRegistry, confirmReadback } from '@kaleidorg/mind';
import { McpToolSource } from '@kaleidorg/mind/mcp';
import { createQvacProvider } from '@kaleidorg/mind/qvac';
@@ -136,11 +136,15 @@ await kaleido.connect();
// 2. A local model through QVAC (downloaded on first run)
const modelId = await loadModel({
- modelSrc: QWEN3_600M_INST_Q4,
+ modelSrc: QWEN3_5_2B_MULTIMODAL_Q4_K_M,
modelType: 'llm',
modelConfig: { ctx_size: 8192, tools: true },
});
-const provider = createQvacProvider({ completion, cancel, getModelId: () => modelId });
+const provider = createQvacProvider({
+ completion, cancel, getModelId: () => modelId,
+ defaultMaxTokens: 1536, // Qwen3.5 reasons before answering; leave room for both
+ maxThinkingTokens: 512,
+});
// 3. The tiered funnel, with a confirmation prompt before any spend
const funnel = new Funnel({ provider, tools: new ToolRegistry([kaleido]) });
@@ -168,7 +172,7 @@ node agent.mjs
The first start downloads the model. Every fund-moving tool pauses on the `onConfirm` callback, so nothing leaves the node until you type `y`.
- This is a minimal host. The [`@kaleidorg/mind` README](https://www.npmjs.com/package/@kaleidorg/mind) and the `examples/node-minimal` and `examples/rgb-agent` folders in the [kaleido-mind repository](https://github.com/kaleidoswap/kaleido-mind) go further, with skills, recipes, and a larger model. A 0.6B model handles the fast path and the recipes well; for open-ended requests a larger model such as Qwen3 1.7B or 4B does noticeably better.
+ This is a minimal host. The [`@kaleidorg/mind` README](https://www.npmjs.com/package/@kaleidorg/mind) and the `examples/node-minimal` and `examples/rgb-agent` folders in the [kaleido-mind repository](https://github.com/kaleidoswap/kaleido-mind) go further, with skills, recipes, and a larger model. The example loads Qwen3.5 2B (about 1.3 GB), which handled every signet RGB wallet task in our tests. `QWEN3_5_4B_MULTIMODAL_Q4_K_M` (about 2.7 GB) is as accurate but about twice as slow; the 0.8B model loops on wallet actions. The token caps in the sample leave room for its reasoning before the answer.
## 5. Prompts to Try
diff --git a/mintlify-docs/ai-tools/faq.mdx b/mintlify-docs/ai-tools/faq.mdx
index 654bc40..594de22 100644
--- a/mintlify-docs/ai-tools/faq.mdx
+++ b/mintlify-docs/ai-tools/faq.mdx
@@ -41,7 +41,7 @@ description: "Common questions about the KaleidoSwap AI tools, covering custody,
It varies by surface:
- **KaleidoAgent** reasons with a hosted model, Claude or OpenAI, selected with `AGENT_PROVIDER`.
- - **KaleidoMind** runs the model **on the device** through the QVAC SDK, locally or on a desktop you explicitly paired. Its tiered funnel means most requests never reach the model at all.
+ - **KaleidoMind** runs the model **on the device** through the QVAC SDK. The recommended model is Qwen3.5 2B, on desktop and phone alike. Its tiered funnel means most requests never reach the model at all. Pairing a phone with the desktop app for delegated inference is paused in desktop app v0.5.1.
- **MCP servers** are model-agnostic. Whatever model your MCP host runs is the one calling the tools.
diff --git a/mintlify-docs/ai-tools/kaleido-mind.mdx b/mintlify-docs/ai-tools/kaleido-mind.mdx
index da091cf..17684f7 100644
--- a/mintlify-docs/ai-tools/kaleido-mind.mdx
+++ b/mintlify-docs/ai-tools/kaleido-mind.mdx
@@ -65,9 +65,23 @@ LLM, embedding, speech-to-text, and text-to-speech inference all run through the
| Package | Version |
|---------|---------|
-| `@qvac/sdk` | The core engine works with 0.13 and later. Install the latest release, which is what the hosts and evals are tested against |
+| `@qvac/sdk` | The core engine works with 0.13.1 and later and is tested with 0.21. P2P delegation needs 0.13.1–0.18: QVAC removed it in 0.19 |
| `@kaleidorg/mind` | Takes `@qvac/sdk` as an optional peer dependency, so the engine also runs with no model at all against the mock provider |
+### Recommended models
+
+KaleidoMind 0.8 recommends the Qwen3.5 family. Every model below is a `@qvac/sdk` registry constant, so `loadModel({ modelSrc: QWEN3_5_2B_MULTIMODAL_Q4_K_M })` downloads and caches it. Qwen3 and Hermes 3 are no longer recommended: the small Qwen3 models garbled the arguments of multi-field wallet calls such as `rln_issue_asset`.
+
+| Model | `@qvac/sdk` constant | Download | RAM | Use |
+|---|---|---|---|---|
+| Qwen3.5 0.8B | `QWEN3_5_0_8B_MULTIMODAL_Q4_K_M` | 0.53 GB | ~1.5 GB | Smoke tests only; loops on wallet actions |
+| Qwen3.5 2B | `QWEN3_5_2B_MULTIMODAL_Q4_K_M` | 1.3 GB | ~3 GB | **Default.** Passed all 7 signet RGB wallet tasks, 45–150 s per question on an M4 laptop |
+| Qwen3.5 4B | `QWEN3_5_4B_MULTIMODAL_Q4_K_M` | 2.7 GB | ~5 GB | Same correctness as 2B, about twice as slow |
+| Qwen3.5 9B | `QWEN3_5_9B_MULTIMODAL_Q4_K_M` | 5.7 GB | ~9 GB | 16 GB machines; slow for interactive chat |
+| Qwen3.6 35B-A3B (MoE) | `QWEN3_6_35B_A3B_MULTIMODAL_Q4_K_M` | 22 GB | ~26 GB | 32 GB+ machines |
+
+The same list is exported as `QWEN35_MODELS` from `@kaleidorg/mind/qvac`.
+
`createQvacProvider` turns the SDK's `completion` into the engine's `LLMProvider`. The SDK functions are injected rather than imported, so the host owns the model lifecycle (`loadModel`, `unloadModel`) and the engine never pulls QVAC into a bundle that does not need it.
## Build With It
@@ -88,7 +102,7 @@ For a step-by-step version on signet, follow [Build a Local RGB Agent](/ai-tools
| Host | Role |
|------|------|
| [Rate](https://github.com/kaleidoswap/Rate) | React Native mobile wallet, with local LLM, speech-to-text, neural text-to-speech, and the hands-free voice loop |
-| [Desktop App](/desktop-app/getting-started/introduction) | Runs the engine as in-app chat through the Tauri sidecar (`apps/provider`), which consumes `kaleido-mcp` as a stdio tool source, and can act as the paired inference peer for a phone |
+| [Desktop App](/desktop-app/getting-started/introduction) | Runs the engine as in-app chat through the Tauri sidecar (`apps/provider`), which consumes `kaleido-mcp` as a stdio tool source, and can act as the paired inference peer for a phone (phone pairing is paused in desktop app v0.5.1) |
The fastest way to try it is the Desktop App. [Download the latest release](https://kaleidoswap.com/downloads) and follow the [installation guide](/desktop-app/getting-started/installation).
diff --git a/mintlify-docs/cn/ai-tools/build-local-rgb-agent.mdx b/mintlify-docs/cn/ai-tools/build-local-rgb-agent.mdx
index c22d447..6af515b 100644
--- a/mintlify-docs/cn/ai-tools/build-local-rgb-agent.mdx
+++ b/mintlify-docs/cn/ai-tools/build-local-rgb-agent.mdx
@@ -108,11 +108,11 @@ npm init -y && npm pkg set type=module
npm install @kaleidorg/mind @qvac/sdk @modelcontextprotocol/sdk
```
-KaleidoMind 支持 `@qvac/sdk` 0.13 及以上版本，建议安装最新版本。把下面的代码保存为 `agent.mjs`：
+KaleidoMind 支持 `@qvac/sdk` 0.13.1 及以上版本，并以 0.21 为测试基准。把下面的代码保存为 `agent.mjs`：
```js agent.mjs
import { createInterface } from 'node:readline/promises';
-import { completion, cancel, loadModel, QWEN3_600M_INST_Q4 } from '@qvac/sdk';
+import { completion, cancel, loadModel, QWEN3_5_2B_MULTIMODAL_Q4_K_M } from '@qvac/sdk';
import { Funnel, ToolRegistry, confirmReadback } from '@kaleidorg/mind';
import { McpToolSource } from '@kaleidorg/mind/mcp';
import { createQvacProvider } from '@kaleidorg/mind/qvac';
@@ -136,11 +136,15 @@ await kaleido.connect();
// 2. 通过 QVAC 加载本地模型（首次运行时下载）
const modelId = await loadModel({
- modelSrc: QWEN3_600M_INST_Q4,
+ modelSrc: QWEN3_5_2B_MULTIMODAL_Q4_K_M,
modelType: 'llm',
modelConfig: { ctx_size: 8192, tools: true },
});
-const provider = createQvacProvider({ completion, cancel, getModelId: () => modelId });
+const provider = createQvacProvider({
+ completion, cancel, getModelId: () => modelId,
+ defaultMaxTokens: 1536, // Qwen3.5 reasons before answering; leave room for both
+ maxThinkingTokens: 512,
+});
// 3. 分层漏斗，任何花费之前都会弹出确认
const funnel = new Funnel({ provider, tools: new ToolRegistry([kaleido]) });
@@ -168,7 +172,7 @@ node agent.mjs
首次启动会下载模型。每个动用资金的工具都会在 `onConfirm` 回调处暂停，在你输入 `y` 之前，任何资金都不会离开节点。
- 这是一个最小的宿主。[`@kaleidorg/mind` README](https://www.npmjs.com/package/@kaleidorg/mind) 以及 [kaleido-mind 仓库](https://github.com/kaleidoswap/kaleido-mind)中的 `examples/node-minimal` 和 `examples/rgb-agent` 目录走得更远，包含 skills、recipe 和更大的模型。0.6B 模型能很好地处理快速路径和 recipe；对于开放式请求，Qwen3 1.7B 或 4B 这类更大的模型表现明显更好。
+ 这是一个最小的宿主。[`@kaleidorg/mind` README](https://www.npmjs.com/package/@kaleidorg/mind) 以及 [kaleido-mind 仓库](https://github.com/kaleidoswap/kaleido-mind)中的 `examples/node-minimal` 和 `examples/rgb-agent` 目录走得更远，包含 skills、recipe 和更大的模型。示例加载 Qwen3.5 2B（约 1.3 GB），在我们的测试中它完成了所有 signet RGB 钱包任务。`QWEN3_5_4B_MULTIMODAL_Q4_K_M`（约 2.7 GB）同样准确，但速度约慢一倍；0.8B 模型在钱包操作上会陷入循环。示例中的 token 上限为模型回答前的推理留出了空间。
## 5. 可以尝试的提示词 {#5-prompts-to-try}
diff --git a/mintlify-docs/cn/ai-tools/faq.mdx b/mintlify-docs/cn/ai-tools/faq.mdx
index 05bf808..403e22e 100644
--- a/mintlify-docs/cn/ai-tools/faq.mdx
+++ b/mintlify-docs/cn/ai-tools/faq.mdx
@@ -41,7 +41,7 @@ description: "关于 KaleidoSwap AI 工具的常见问题：托管方式、种
因接入方式而异：
- **KaleidoAgent** 用托管模型推理，Claude 或 OpenAI，通过 `AGENT_PROVIDER` 选择。
- - **KaleidoMind** 通过 QVAC SDK 让模型**在设备上**运行，可以在本地，也可以在你显式配对过的桌面端。它的分层漏斗设计意味着大多数请求根本不会到达模型。
+ - **KaleidoMind** 通过 QVAC SDK 让模型**在设备上**运行。推荐使用 Qwen3.5 2B，桌面端和手机都适用。它的分层漏斗设计意味着大多数请求根本不会到达模型。桌面应用 v0.5.1 中，手机与桌面配对进行委托推理的功能已暂停。
- **MCP 服务器** 与模型无关。调用工具的就是你的 MCP 宿主所运行的那个模型。
diff --git a/mintlify-docs/cn/ai-tools/kaleido-mind.mdx b/mintlify-docs/cn/ai-tools/kaleido-mind.mdx
index c175927..64fe52e 100644
--- a/mintlify-docs/cn/ai-tools/kaleido-mind.mdx
+++ b/mintlify-docs/cn/ai-tools/kaleido-mind.mdx
@@ -65,9 +65,23 @@ LLM、嵌入、语音转文字和文字转语音推理全部通过 Tether 的本
| 包 | 版本 |
|---------|---------|
-| `@qvac/sdk` | 核心引擎支持 0.13 及以上版本。建议安装最新版本，宿主与评测都以它为测试基准 |
+| `@qvac/sdk` | 核心引擎支持 0.13.1 及以上版本，并以 0.21 为测试基准。P2P 委托需要 0.13.1–0.18：QVAC 在 0.19 中移除了该功能 |
| `@kaleidorg/mind` | 把 `@qvac/sdk` 作为可选 peer 依赖，因此引擎在完全没有模型的情况下也能配合 mock provider 运行 |
+### 推荐模型
+
+KaleidoMind 0.8 推荐 Qwen3.5 系列。下表中的每个模型都是 `@qvac/sdk` 注册表常量，`loadModel({ modelSrc: QWEN3_5_2B_MULTIMODAL_Q4_K_M })` 会自动下载并缓存。不再推荐 Qwen3 和 Hermes 3：较小的 Qwen3 模型会弄乱 `rln_issue_asset` 等多字段钱包调用的参数。
+
+| 模型 | `@qvac/sdk` 常量 | 下载大小 | 内存 | 用途 |
+|---|---|---|---|---|
+| Qwen3.5 0.8B | `QWEN3_5_0_8B_MULTIMODAL_Q4_K_M` | 0.53 GB | 约 1.5 GB | 仅用于冒烟测试；执行钱包操作时会陷入循环 |
+| Qwen3.5 2B | `QWEN3_5_2B_MULTIMODAL_Q4_K_M` | 1.3 GB | 约 3 GB | **默认。** 在 M4 笔记本上通过全部 7 项 signet RGB 钱包任务，每个问题 45–150 秒 |
+| Qwen3.5 4B | `QWEN3_5_4B_MULTIMODAL_Q4_K_M` | 2.7 GB | 约 5 GB | 正确率与 2B 相同，速度约慢一倍 |
+| Qwen3.5 9B | `QWEN3_5_9B_MULTIMODAL_Q4_K_M` | 5.7 GB | 约 9 GB | 16 GB 内存的机器；交互式聊天较慢 |
+| Qwen3.6 35B-A3B（MoE） | `QWEN3_6_35B_A3B_MULTIMODAL_Q4_K_M` | 22 GB | 约 26 GB | 32 GB 以上内存的机器 |
+
+同一份列表也以 `QWEN35_MODELS` 从 `@kaleidorg/mind/qvac` 导出。
+
`createQvacProvider` 把 SDK 的 `completion` 转换为引擎所需的 `LLMProvider`。SDK 函数通过注入而非导入的方式传入，因此由宿主掌控模型生命周期（`loadModel`、`unloadModel`），引擎也不会把 QVAC 打进不需要它的包里。
## 动手构建 {#build-with-it}
@@ -88,7 +102,7 @@ LLM、嵌入、语音转文字和文字转语音推理全部通过 Tether 的本
| 宿主 | 角色 |
|------|------|
| [Rate](https://github.com/kaleidoswap/Rate) | React Native 移动钱包，内含本地 LLM、语音转文字、神经网络文字转语音，以及免手操作的语音循环 |
-| [桌面应用](/cn/desktop-app/getting-started/introduction) | 通过 Tauri sidecar（`apps/provider`）把引擎作为应用内聊天运行，该 sidecar 以 stdio 工具源方式接入 `kaleido-mcp`，同时还能作为手机的配对推理节点 |
+| [桌面应用](/cn/desktop-app/getting-started/introduction) | 通过 Tauri sidecar（`apps/provider`）把引擎作为应用内聊天运行，该 sidecar 以 stdio 工具源方式接入 `kaleido-mcp`，同时还能作为手机的配对推理节点（桌面应用 v0.5.1 中手机配对功能暂停） |
最快的体验方式是桌面应用。[下载最新版本](https://kaleidoswap.com/downloads)，然后按照[安装指南](/cn/desktop-app/getting-started/installation)操作。