Skip to content

A local-first multi-agent app development platform, built on PenguinHarness

feat(core): 模型页多了小米、百度、腾讯、讯飞四家的预置模型

lmliheng650bd49e9da2
diff --git a/changelog/unreleased/2026-10-15-china-model-vendors.md b/changelog/unreleased/2026-10-15-china-model-vendors.md
new file mode 100644
index 00000000..1d3d2214
--- /dev/null
+++ b/changelog/unreleased/2026-10-15-china-model-vendors.md
@@ -0,0 +1,32 @@
+# Four more first-party model vendors: Xiaomi, Baidu, Tencent and iFlytek
+
+- **Date:** 2026-10-15
+- **Type:** feat
+- **Scope:** `core`
+
+[中文版](2026-10-15-china-model-vendors.zh.md)
+
+The Models page gains four preset groups, each carrying that vendor's own models and endpoint, so
+their models can be used by pasting a key instead of filling in an endpoint by hand:
+
+| Group | Models | Endpoint |
+| --- | --- | --- |
+| Xiaomi MiMo | `mimo-v2.6-pro`, `mimo-v2.6-flash` (both 1M context, image input) | `https://api.xiaomimimo.com/v1` |
+| Baidu ERNIE | `ernie-5.1`, `ernie-5.0`, `ernie-4.5-turbo-vl`, `ernie-4.5-turbo-128k`, `ernie-x1.1` | `https://qianfan.baidubce.com/v2` (Qianfan's OpenAI-compatible base is `/v2`; `/v1` is retired) |
+| Tencent Hunyuan | `hunyuan-a13b`, `hunyuan-vision-1.5-instruct`, `hunyuan-t1-vision-20250916`, `hunyuan-turbos-vision-video` | `https://api.hunyuan.cloud.tencent.com/v1` |
+| iFlytek Spark | `4.0Ultra`, `pro-128k`, `lite` | `https://spark-api-open.xf-yun.com/v1` |
+
+Every group links its own console (where the key is created) and its model list, and every row
+records what the vendor publishes: context window, image input, and the price converted from the
+vendor's CNY list at the catalog's usual 7:1 rate. AgentHub has no first-party client for any of
+the four, so each row pins the Chat Completions protocol and inlines the endpoint — the same shape
+the ModelScope and vLLM rows already use.
+
+Where the vendors' own documentation is thin, the catalog says so instead of filling the gap:
+iFlytek's rows carry no price (`4.0Ultra`, `pro-128k`) or the free rate it publishes (`lite`),
+because its HTTP doc has no price list and its product page prices model families it has published
+no API id for; Spark's Max pair is left out (Spark dates those plans offline on 2026-03-10).
+Tencent is migrating this platform to TokenHub, so its console link is the one today's docs name,
+and the role-play / translation / embedding rows are left out as single-purpose. Xiaomi's
+`mimo-v2.6-pro-ultraspeed` is left out too: its model page carries no capability column, so neither
+a window nor vision support could be read from it.
diff --git a/changelog/unreleased/2026-10-15-china-model-vendors.zh.md b/changelog/unreleased/2026-10-15-china-model-vendors.zh.md
new file mode 100644
index 00000000..a6ae453c
--- /dev/null
+++ b/changelog/unreleased/2026-10-15-china-model-vendors.zh.md
@@ -0,0 +1,26 @@
+# 新增四家国内模型商:小米、百度、腾讯、讯飞
+
+- **Date:** 2026-10-15
+- **Type:** feat
+- **Scope:** `core`
+
+[English](2026-10-15-china-model-vendors.md)
+
+模型页多了四个预置分组,每组都带着厂商自己的模型与端点——贴一把 key 就能用,不必再手填 base URL:
+
+| 分组 | 模型 | 端点 |
+| --- | --- | --- |
+| 小米 MiMo | `mimo-v2.6-pro`、`mimo-v2.6-flash`(均 1M 上下文、支持图片输入) | `https://api.xiaomimimo.com/v1` |
+| 百度文心 ERNIE | `ernie-5.1`、`ernie-5.0`、`ernie-4.5-turbo-vl`、`ernie-4.5-turbo-128k`、`ernie-x1.1` | `https://qianfan.baidubce.com/v2`(千帆的 OpenAI 兼容路径是 `/v2`,`/v1` 已下线) |
+| 腾讯混元 | `hunyuan-a13b`、`hunyuan-vision-1.5-instruct`、`hunyuan-t1-vision-20250916`、`hunyuan-turbos-vision-video` | `https://api.hunyuan.cloud.tencent.com/v1` |
+| 讯飞星火 | `4.0Ultra`、`pro-128k`、`lite` | `https://spark-api-open.xf-yun.com/v1` |
+
+每个分组都链到各自的控制台(建 key 的那一页)与模型清单;每一行记下厂商公布的东西——上下文窗口、
+是否吃图片、以及按目录惯例 7:1 折算成美元的价格(厂商原本是人民币)。这四家 AgentHub 都没有一方
+客户端,所以每行都钉住 Chat Completions 协议并内联端点——与 ModelScope、vLLM 现有的写法一致。
+
+厂商文档本身讲不清的地方,目录里如实写出来、而不是替它补齐:讯飞那几行不带价格(`4.0Ultra`、
+`pro-128k`)或用它自己公布的免费价(`lite`)——它的 HTTP 文档没有价格表,产品页的价格又只按它没公布
+过 API id 的 X2.5 系列列出;星火的 Max 一对也没收录(官方写明这两个套餐 2026-03-10 下线)。腾讯这条
+线正在迁往 TokenHub,控制台链接取的是它今天文档里那一个;角色扮演 / 翻译 / 向量化三类偏单用途的行没进
+预置。小米的 `mimo-v2.6-pro-ultraspeed` 同样没收:它的模型页没有能力列,窗口与图片支持都读不出来。
diff --git a/packages/core/src/state/model-catalog.ts b/packages/core/src/state/model-catalog.ts
index 50bfeaad..89e96046 100644
--- a/packages/core/src/state/model-catalog.ts
+++ b/packages/core/src/state/model-catalog.ts
@@ -23,7 +23,11 @@
* from public model pages and endpoint listings — their windows, vision flags and prices are
* NOT verified and say so on the rows themselves: 2026-09-18 and 2026-09-20; and the direct
* Z.AI glm-4.7-flash row, the group's free tier, read from Z.AI's price list and
- * 智谱开放平台's free-model page: 2026-10-07).
+ * 智谱开放平台's free-model page: 2026-10-07; and the four Chinese first-party groups added
+ * together — Xiaomi MiMo, Baidu ERNIE, Tencent Hunyuan and iFlytek Spark — read from each
+ * vendor's own docs and console pages: 2026-10-15, with the iFlytek rows deliberately unpriced
+ * (its HTTP doc publishes no price list, and the product page's prices name model families it
+ * has published no API id for).
* Docs: packages/docs/content/models.{zh,en}.md (site path /docs/models) documents the
* provider groups and credential resolution described here.
*
@@ -282,6 +286,14 @@ const TOKENDANCE_BASE_URL = "https://tokendance.space/gateway/v1";
const MINIMAX_BASE_URL = "https://api.minimax.io/v1";
const DEEPSEEK_BASE_URL = "https://api.deepseek.com";
const MODELSCOPE_BASE_URL = "https://api-inference.modelscope.cn/v1";
+/** Xiaomi MiMo's pay-as-you-go endpoint (its batch and token-plan hosts are separate services with their own keys). */
+const XIAOMI_BASE_URL = "https://api.xiaomimimo.com/v1";
+/** Baidu Qianfan's OpenAI-compatible base: `/v2` — `/v1` is the retired interface. */
+const BAIDU_BASE_URL = "https://qianfan.baidubce.com/v2";
+/** Tencent Hunyuan's OpenAI-compatible endpoint. */
+const HUNYUAN_BASE_URL = "https://api.hunyuan.cloud.tencent.com/v1";
+/** iFlytek Spark's OpenAI-compatible endpoint (the WebSocket API is a different, signed protocol). */
+const IFLYTEK_BASE_URL = "https://spark-api-open.xf-yun.com/v1";
export const PENGUIN_GO_BASE_URL = "https://token.penguin.ooo/api";
/** OpenCode Go's OpenAI-protocol base: the clients append /chat/completions or /responses. */
const OPENCODE_GO_BASE_URL = "https://opencode.ai/zen/go/v1";
@@ -453,6 +465,66 @@ export const MODEL_PROVIDERS: ModelProviderInfo[] = [
apiKeyUrl: "https://platform.minimax.io/user-center/basic-information/interface-key",
modelsUrl: "https://platform.minimax.io/docs/guides/models-intro",
},
+ {
+ // Xiaomi's own model platform (MiMo API Open Platform). OpenAI-compatible at the
+ // pay-as-you-go endpoint the preset rows inline (the same host's Anthropic-protocol path is
+ // not catalogued), keys from its console, so no oauth or bridgeAuth. AgentHub has no Xiaomi
+ // client, so every row pins `openai-chat` and names the endpoint; the group pins the same
+ // protocol, which is also what keeps it out of the OPENAI_* env fallback
+ // (providerEnvFallbackKey) — the pair below records what that generic client reads when a
+ // row is handed no endpoint, not a variable of Xiaomi's.
+ // https://mimo.mi.com/docs/quick-start/first-api-call, https://mimo.mi.com/docs/price/pay-as-you-go
+ id: "xiaomi",
+ label: "Xiaomi MiMo",
+ envKey: "OPENAI_API_KEY",
+ envBaseUrlKey: "OPENAI_BASE_URL",
+ apiKeyUrl: "https://platform.xiaomimimo.com/#/console/api-keys",
+ modelsUrl: "https://mimo.mi.com/docs/quick-start/summary/model",
+ clientType: "openai-chat",
+ },
+ {
+ // Baidu's Qianfan platform, serving the ERNIE line. Its OpenAI-compatible base is `/v2` —
+ // `/v1` is the retired interface, and nothing here derives the path from a `v1` habit.
+ // https://cloud.baidu.com/doc/qianfan-api/s/3m7of64lb (endpoint and auth),
+ // https://cloud.baidu.com/doc/qianfan/s/rmh4stp0j (model list), .../s/wmh4sv6ya (prices)
+ id: "baidu",
+ label: "Baidu ERNIE",
+ envKey: "OPENAI_API_KEY",
+ envBaseUrlKey: "OPENAI_BASE_URL",
+ apiKeyUrl: "https://console.bce.baidu.com/iam/#/iam/apikey/list",
+ modelsUrl: "https://cloud.baidu.com/doc/qianfan/s/rmh4stp0j",
+ clientType: "openai-chat",
+ },
+ {
+ // Tencent Hunyuan's own endpoint (the OpenAI-compatible one; Tencent also publishes an
+ // Anthropic-compatible path and the TC3-native API, neither catalogued). Tencent is moving
+ // this platform to TokenHub, so the console link below is the one its docs name today.
+ // https://cloud.tencent.com/document/product/1729/111007 (endpoint and auth),
+ // .../1729/104753 (model list), .../1729/97731 (prices)
+ id: "tencent",
+ label: "Tencent Hunyuan",
+ envKey: "OPENAI_API_KEY",
+ envBaseUrlKey: "OPENAI_BASE_URL",
+ apiKeyUrl: "https://console.cloud.tencent.com/hunyuan/start",
+ modelsUrl: "https://cloud.tencent.com/document/product/1729/104753",
+ clientType: "openai-chat",
+ },
+ {
+ // iFlytek Spark (讯飞星火). The OpenAI-compatible endpoint takes the console's APIPassword
+ // as a Bearer token; only the WebSocket API needs the appid/APIKey/APISecret signature, and
+ // nothing here uses it. The documented `model` values are still the 4.0 Ultra / Pro /
+ // Pro-128K / Lite names (the doc names no id for the newer X2.5 and X2 line-ups), and the
+ // doc's own model table dates from 2025 — see the rows in MODEL_CATALOG for what that means
+ // for their prices.
+ // https://www.xfyun.cn/doc/spark/HTTP%E8%B0%83%E7%94%A8%E6%96%87%E6%A1%A3.html
+ id: "iflytek",
+ label: "iFlytek Spark",
+ envKey: "OPENAI_API_KEY",
+ envBaseUrlKey: "OPENAI_BASE_URL",
+ apiKeyUrl: "https://console.xfyun.cn/services/cbm",
+ modelsUrl: "https://www.xfyun.cn/doc/spark/HTTP%E8%B0%83%E7%94%A8%E6%96%87%E6%A1%A3.html",
+ clientType: "openai-chat",
+ },
{
id: "qwen-pay-as-you-go",
label: "Qwen Pay-As-You-Go",
@@ -2882,6 +2954,173 @@ export const MODEL_CATALOG: ModelCatalogEntry[] = [
supportsVision: true,
clientType: VLLM_CLIENT_TYPE,
},
+ // -- Xiaomi MiMo (first-party vendor: OpenAI-compatible, so every row pins the protocol and
+ // inlines the endpoint — AgentHub routes a bare `mimo-*` id nowhere). Ids, windows, vision
+ // flags and prices read off the MiMo open platform's model and price pages on 2026-10-15;
+ // its prices are CNY per million tokens (cache hit / input miss / output), which is the order
+ // `cny` takes. Left out: `mimo-v2.6-pro-ultraspeed` (the model page carries no capability
+ // column for it, so neither a window nor vision support could be read off it) and the v2.5
+ // pair the platform dates for retirement on 2026-10-21.
+ {
+ modelId: "mimo-v2.6-pro",
+ displayName: "MiMo V2.6 Pro",
+ provider: "xiaomi",
+ contextWindow: 1000000,
+ pricing: cny(0.025, 3, 6),
+ supportsVision: true,
+ clientType: "openai-chat",
+ baseUrl: XIAOMI_BASE_URL,
+ },
+ {
+ modelId: "mimo-v2.6-flash",
+ displayName: "MiMo V2.6 Flash",
+ provider: "xiaomi",
+ contextWindow: 1000000,
+ pricing: cny(0.02, 1, 2),
+ supportsVision: true,
+ clientType: "openai-chat",
+ baseUrl: XIAOMI_BASE_URL,
+ },
+ // -- Baidu Qianfan (ERNIE). Prices are the vendor's CNY per thousand tokens scaled to the
+ // million the catalog stores, read on 2026-10-15; the ≥32K-input tier of the ERNIE 5 rows
+ // (6 / 22 CNY) is not recorded, for the same single-rate reason the OpenAI rows record their
+ // base tier. `ernie-5.1` sits in the model list's text-generation section rather than the
+ // vision one, so its row is text-only. Qianfan also resells other vendors' models
+ // (DeepSeek, Qwen, GLM, Kimi); they are left out — each has its own group here.
+ {
+ modelId: "ernie-5.1",
+ displayName: "ERNIE 5.1",
+ provider: "baidu",
+ contextWindow: 131072,
+ pricing: cny(4, 4, 18),
+ supportsVision: false,
+ clientType: "openai-chat",
+ baseUrl: BAIDU_BASE_URL,
+ },
+ {
+ modelId: "ernie-5.0",
+ displayName: "ERNIE 5.0",
+ provider: "baidu",
+ contextWindow: 131072,
+ pricing: cny(6, 6, 24),
+ supportsVision: true,
+ clientType: "openai-chat",
+ baseUrl: BAIDU_BASE_URL,
+ },
+ {
+ modelId: "ernie-4.5-turbo-vl",
+ displayName: "ERNIE 4.5 Turbo VL",
+ provider: "baidu",
+ contextWindow: 131072,
+ pricing: cny(3, 3, 9),
+ supportsVision: true,
+ clientType: "openai-chat",
+ baseUrl: BAIDU_BASE_URL,
+ },
+ {
+ modelId: "ernie-4.5-turbo-128k",
+ displayName: "ERNIE 4.5 Turbo 128K",
+ provider: "baidu",
+ contextWindow: 131072,
+ // The one ERNIE row whose price list publishes a cache-hit rate (0.2 CNY); the others read
+ // their input price, as every row without a published cache price does.
+ pricing: cny(0.2, 0.8, 3.2),
+ supportsVision: false,
+ clientType: "openai-chat",
+ baseUrl: BAIDU_BASE_URL,
+ },
+ {
+ modelId: "ernie-x1.1",
+ displayName: "ERNIE X1.1",
+ provider: "baidu",
+ contextWindow: 65536,
+ pricing: cny(1, 1, 4),
+ supportsVision: false,
+ clientType: "openai-chat",
+ baseUrl: BAIDU_BASE_URL,
+ },
+ // -- Tencent Hunyuan. The windows are the model list's "maximum input" (Hunyuan publishes a
+ // separate maximum output beside it, which the catalog's single window cannot carry). Prices
+ // are the vendor's CNY per million tokens, read on 2026-10-15. Left out: the role-play and
+ // translation rows (single-purpose), the embedding row (not a chat model), and
+ // `hunyuan-turbos-latest` / `hunyuan-vision`, which the compatibility example names but the
+ // model list does not describe.
+ {
+ modelId: "hunyuan-a13b",
+ displayName: "Hunyuan A13B",
+ provider: "tencent",
+ contextWindow: 229376,
+ pricing: cny(0.5, 0.5, 2),
+ supportsVision: false,
+ clientType: "openai-chat",
+ baseUrl: HUNYUAN_BASE_URL,
+ },
+ {
+ modelId: "hunyuan-vision-1.5-instruct",
+ displayName: "Tencent HY Vision 1.5 Instruct",
+ provider: "tencent",
+ contextWindow: 24576,
+ pricing: cny(3, 3, 9),
+ supportsVision: true,
+ clientType: "openai-chat",
+ baseUrl: HUNYUAN_BASE_URL,
+ },
+ {
+ modelId: "hunyuan-t1-vision-20250916",
+ displayName: "Hunyuan T1 Vision",
+ provider: "tencent",
+ contextWindow: 28672,
+ pricing: cny(3, 3, 9),
+ supportsVision: true,
+ clientType: "openai-chat",
+ baseUrl: HUNYUAN_BASE_URL,
+ },
+ {
+ modelId: "hunyuan-turbos-vision-video",
+ displayName: "Hunyuan TurboS Vision Video",
+ provider: "tencent",
+ contextWindow: 24576,
+ pricing: cny(3, 3, 9),
+ supportsVision: true,
+ clientType: "openai-chat",
+ baseUrl: HUNYUAN_BASE_URL,
+ },
+ // -- iFlytek Spark. The `model` values and the windows are the HTTP doc's own table (Ultra
+ // 32K in / 32K out, Pro-128K 128K, Lite 8K; the Max pair the table also lists is left out —
+ // Spark states its Max plans went offline on 2026-03-10 and their traffic moved to Ultra).
+ // That doc publishes no price list, and the product page's prices name X2.5 / X2 model
+ // families for which iFlytek has published no API id, so those rows carry no price at all
+ // rather than a guessed or borrowed one; Lite is the one the doc calls free, and its zero is
+ // the vendor's. Both facts are why this group is the catalog's weakest: re-read it before
+ // trusting a row.
+ {
+ modelId: "4.0Ultra",
+ displayName: "Spark 4.0 Ultra",
+ provider: "iflytek",
+ contextWindow: 32768,
+ supportsVision: false,
+ clientType: "openai-chat",
+ baseUrl: IFLYTEK_BASE_URL,
+ },
+ {
+ modelId: "pro-128k",
+ displayName: "Spark Pro-128K",
+ provider: "iflytek",
+ contextWindow: 131072,
+ supportsVision: false,
+ clientType: "openai-chat",
+ baseUrl: IFLYTEK_BASE_URL,
+ },
+ {
+ modelId: "lite",
+ displayName: "Spark Lite",
+ provider: "iflytek",
+ contextWindow: 8192,
+ pricing: usd(0, 0, 0),
+ supportsVision: false,
+ clientType: "openai-chat",
+ baseUrl: IFLYTEK_BASE_URL,
+ },
// -- Custom (the group that otherwise holds only user-defined models). A preset may live
// here only as a complete row — its own base URL and a pinned generic protocol — because the
// group implies neither; it is where a vendor with a single preview model and no console
diff --git a/packages/core/test/model-catalog.test.ts b/packages/core/test/model-catalog.test.ts
index 03a4598b..5e84dc30 100644
--- a/packages/core/test/model-catalog.test.ts
+++ b/packages/core/test/model-catalog.test.ts
@@ -68,6 +68,10 @@ describe("model-catalog", () => {
"zhipu",
"moonshot",
"minimax",
+ "xiaomi",
+ "baidu",
+ "tencent",
+ "iflytek",
"qwen-pay-as-you-go",
"qwen-token-plan",
"modelscope",
@@ -380,6 +384,11 @@ describe("model-catalog", () => {
"deepseek-ai/DeepSeek-V4.1-Flash",
"Qwen/Qwen3.8-27B",
"Qwen/Qwen3.8-Flash-Next",
+ // iFlytek Spark's HTTP doc publishes no price list at all, and the prices on its product
+ // page name the X2.5 / X2 families, for which it has published no API id — so these rows
+ // are recorded as unpriced rather than priced from a neighbouring model.
+ "4.0Ultra",
+ "pro-128k",
]);
for (const m of MODEL_CATALOG) {
if (m.pricing === undefined) {
@@ -390,7 +399,8 @@ describe("model-catalog", () => {
m.modelId === "openrouter/free" ||
m.modelId === "Atria-Dawn-Preview" ||
m.modelId === "dots-3-note-preview" ||
- m.modelId === "glm-4.7-flash"
+ m.modelId === "glm-4.7-flash" ||
+ m.modelId === "lite"
) {
// Self-hosted vLLM and the free-tier gateway rows share one treatment: a genuine $0
// price (not "unknown"), so costs compute to 0 and the free badge shows. Nobody bills
@@ -486,12 +496,18 @@ describe("model-catalog", () => {
// The group pin, read the one way every call site reads it.
expect(providerClientType("vllm")).toBe("openai-chat-vllm-adapter");
expect(providerInfo("vllm")!.gatewayBaseUrl).toBeUndefined();
- // Two groups declare one, in MODEL_PROVIDERS order: OpenRouter, whose entries all speak
- // the Responses API its preset endpoint serves, and vLLM. The remaining gateways derive
- // openai-chat from their preset endpoint, and custom / user-defined groups leave the
- // protocol to detection.
+ // Groups that declare one at group level, in MODEL_PROVIDERS order: OpenRouter, whose
+ // entries all speak the Responses API its preset endpoint serves; vLLM; and the four
+ // Chinese first-party groups added 2026-10-15 (Xiaomi MiMo, Baidu ERNIE, Tencent Hunyuan,
+ // iFlytek Spark), whose rows all speak Chat Completions to an endpoint AgentHub has no
+ // first-party client for. The remaining gateways derive openai-chat from their preset
+ // endpoint, and custom / user-defined groups leave the protocol to detection.
expect(MODEL_PROVIDERS.filter((p) => p.clientType !== undefined).map((p) => p.id)).toEqual([
"openrouter",
+ "xiaomi",
+ "baidu",
+ "tencent",
+ "iflytek",
"vllm",
]);
expect(providerClientType("openrouter")).toBe("openai-responses");
@@ -969,6 +985,10 @@ describe("model-catalog", () => {
"modelscope",
"vllm",
"custom",
+ "xiaomi",
+ "baidu",
+ "tencent",
+ "iflytek",
]) {
expect(providerInfo(id)!.envKey).toBe("OPENAI_API_KEY");
expect(providerInfo(id)!.envBaseUrlKey).toBe("OPENAI_BASE_URL");
@@ -1037,7 +1057,8 @@ describe("model-catalog", () => {
// In preset entries, every gateway model inlines base_url, and so do the two direct rows
// whose own id does not route — MiniMax M3 and DeepSeek deepseek-flash (no credentials) —
- // and the custom group's preset, whose group implies no endpoint at all.
+ // the custom group's preset, whose group implies no endpoint at all, and the four Chinese
+ // first-party groups added 2026-10-15, whose endpoint AgentHub has no client for.
const pinnedDirect = MODEL_CATALOG.filter((m) => m.provider === "deepseek" && m.baseUrl);
expect(pinnedDirect.map((m) => m.modelId)).toEqual(["deepseek-flash"]);
const customPresets = MODEL_CATALOG.filter((m) => m.provider === "custom");
@@ -1050,6 +1071,9 @@ describe("model-catalog", () => {
...customPresets,
...MODEL_CATALOG.filter((m) => m.provider === "penguin-go"),
...MODEL_CATALOG.filter((m) => m.provider === "opencode-go"),
+ ...MODEL_CATALOG.filter((m) =>
+ ["xiaomi", "baidu", "tencent", "iflytek"].includes(m.provider),
+ ),
]
// A retired gateway row keeps its pin in the catalog but is not a preset.
.filter((m) => m.retired !== true)
diff --git a/packages/ui/src/components/icons/logos/provider-logo.tsx b/packages/ui/src/components/icons/logos/provider-logo.tsx
index 22af1c56..7f097d59 100644
--- a/packages/ui/src/components/icons/logos/provider-logo.tsx
+++ b/packages/ui/src/components/icons/logos/provider-logo.tsx
@@ -20,6 +20,12 @@
* colored ink on a light tint hashed from the id (the soft tinted-tile style
* AgentAvatar uses, keeping the row's visual weight close to the flat preset
* glyphs).
+ *
+ * The Chinese first-party groups added with their models (Xiaomi MiMo, Baidu
+ * ERNIE, Tencent Hunyuan, iFlytek Spark) have no glyph here either: no brand
+ * asset for them was verified, and a hand-drawn approximation of a trademark
+ * is worse than the tile. They therefore take the same letter tile as a
+ * user-defined group until someone adds a mark they can point at.
*/
import type { ReactNode } from "react";
import { avatarInitial, avatarTile } from "../avatars/avatar";
diff --git a/packages/web/test/model-grouping.test.ts b/packages/web/test/model-grouping.test.ts
index 340d5efe..1fcb6148 100644
--- a/packages/web/test/model-grouping.test.ts
+++ b/packages/web/test/model-grouping.test.ts
@@ -105,6 +105,10 @@ describe("groupModelRows", () => {
"zhipu",
"moonshot",
"minimax",
+ "xiaomi",
+ "baidu",
+ "tencent",
+ "iflytek",
"qwen-pay-as-you-go",
"qwen-token-plan",
"modelscope",