Skip to content

Local UI to run and train LLMs and diffusion models. Supports GGUF, MLX, Qwen3.8, DeepSeek-V4.1, MiniMax-H3, Gemma 4, FLUX and more.

Studio: add QwenCloud as an API provider (#14033)

Nilayf172c8d7c4bc
diff --git a/studio/backend/core/inference/external_provider.py b/studio/backend/core/inference/external_provider.py
index fc87e702568..5de4f935ae5 100644
--- a/studio/backend/core/inference/external_provider.py
+++ b/studio/backend/core/inference/external_provider.py
@@ -46,7 +46,7 @@ _CONTINUATION_FLAG_PROVIDERS = frozenset({"vllm", "llama_cpp"})
# custom may reject include_usage; OpenAI Responses returns usage without it; listed streams need it.
_USAGE_STREAM_OPTION_PROVIDERS = frozenset(
- {"vllm", "llama_cpp", "ollama", "openrouter", "kimi", "lemonade", "qwen"}
+ {"vllm", "llama_cpp", "ollama", "openrouter", "kimi", "lemonade", "qwen", "qwencloud"}
)
# launch-time windows are absent from catalogues; custom covers unregistered self-hosted servers.
@@ -1859,7 +1859,7 @@ class ExternalProviderClient:
_apply_mistral_reasoning_controls(body, model, enable_thinking, reasoning_effort)
elif self.provider_type == "deepseek":
_apply_deepseek_reasoning_controls(body, enable_thinking, reasoning_effort)
- elif self.provider_type == "qwen":
+ elif self.provider_type in ("qwen", "qwencloud"):
_apply_qwen_reasoning_controls(body, enable_thinking)
elif self.provider_type == "huggingface":
_apply_passthrough_reasoning_effort(body, enable_thinking, reasoning_effort)
diff --git a/studio/backend/core/inference/providers.py b/studio/backend/core/inference/providers.py
index 752cc108f2e..f98fdc38812 100644
--- a/studio/backend/core/inference/providers.py
+++ b/studio/backend/core/inference/providers.py
@@ -297,6 +297,32 @@ PROVIDER_REGISTRY: dict[str, dict[str, Any]] = {
"auth_prefix": "Bearer ",
"notes": "DashScope API key. China mainland: override base URL to https://dashscope.aliyuncs.com/compatible-mode/v1",
},
+ "qwencloud": {
+ "display_name": "QwenCloud",
+ "base_url": "https://maas.qwencloudapi.com/compatible-mode/v1",
+ "default_models": [
+ "qwen3.8-max",
+ "qwen3.7-plus",
+ "qwen3.8-flash",
+ ],
+ "model_capabilities": {
+ "qwen3.8-max": {"vision": True, "studio_tools": True},
+ "qwen3.7-plus": {"vision": True, "studio_tools": True},
+ "qwen3.8-flash": {"vision": True, "studio_tools": True},
+ },
+ "supports_streaming": True,
+ "supports_vision": True,
+ "supports_tool_calling": True,
+ "studio_tools": True,
+ "auth_header": "Authorization",
+ "auth_prefix": "Bearer ",
+ "notes": "QwenCloud pay-as-you-go API key (sk-ws-) from https://home.qwencloud.com/api-keys",
+ "model_id_denylist": re.compile(
+ r"(?:^|-)(?:realtime|livetranslate|tts|asr|s2s|embedding|rerank|ocr)\b"
+ r"|^(?:qwen-audio|qwen-image|wan\d|z-image|cosyvoice|happyhorse|happyoyster|"
+ r"tongyi-tingwu|ccai-pro|pre-mock)"
+ ),
+ },
"huggingface": {
"display_name": "Hugging Face",
"base_url": "https://router.huggingface.co/v1",
diff --git a/studio/backend/core/research_runs.py b/studio/backend/core/research_runs.py
index 178cbcb1336..33c93543c73 100644
--- a/studio/backend/core/research_runs.py
+++ b/studio/backend/core/research_runs.py
@@ -1766,7 +1766,13 @@ class ResearchSupervisor:
)
if not omit_sampling and inference.get("topP") is not None:
payload["top_p"] = inference["topP"]
- if inference.get("providerType") in ("deepseek", "huggingface", "qwen", "mistral"):
+ if inference.get("providerType") in (
+ "deepseek",
+ "huggingface",
+ "qwen",
+ "qwencloud",
+ "mistral",
+ ):
# These providers forward reasoning fields verbatim, and a strict upstream rejects one the model lacks.
# Mistral documents reasoning_effort for mistral-small-latest and mistral-medium-3-5 only, so the
# planner opt-out must not reach mistral-large and the other non-reasoning models.
diff --git a/studio/backend/routes/inference.py b/studio/backend/routes/inference.py
index 8ed88aabfbd..f2cd2d1d53e 100644
--- a/studio/backend/routes/inference.py
+++ b/studio/backend/routes/inference.py
@@ -10096,7 +10096,11 @@ _TEXT_ONLY_PROVIDER_TYPES = frozenset({"cohere", "deepseek", "mistral"})
# because its allowlist permits only multimodal kimi-k3/k2.5/k2.6.
# Single-endpoint types (custom, vllm, ollama, llamacpp, lmstudio) retain the
# registry default: no model catalog exists to check the configured endpoint.
-_MIXED_CATALOG_PROVIDER_TYPES = frozenset({"huggingface", "openrouter", "qwen"})
+_MIXED_CATALOG_PROVIDER_TYPES = frozenset({"huggingface", "openrouter", "qwen", "qwencloud"})
+
+# Mixed catalogs whose models.dev namespace answers vision per model.
+# Mirrors CATALOG_VISION_PROVIDER_TYPES in external-providers.ts.
+_CATALOG_VISION_NAMESPACES = {"qwencloud": "qwen"}
def _external_takes_mcp_images(
@@ -10120,6 +10124,12 @@ def _external_takes_mcp_images(
if model and model in capabilities:
return bool(capabilities[model].get("vision"))
if provider_type in _MIXED_CATALOG_PROVIDER_TYPES:
+ namespace = _CATALOG_VISION_NAMESPACES.get(provider_type)
+ if namespace and model:
+ from routes.providers import cached_catalog_model_inputs
+ inputs = cached_catalog_model_inputs(namespace, model)
+ if inputs is not None:
+ return "image" in inputs
# Vision for the family, unknown for this model: an image would be sent to
# something that may not take one, and the failure lands mid-run.
return False
diff --git a/studio/backend/routes/providers.py b/studio/backend/routes/providers.py
index 768406818a0..6d920fd53c1 100644
--- a/studio/backend/routes/providers.py
+++ b/studio/backend/routes/providers.py
@@ -949,6 +949,16 @@ def _write_model_catalog_file(data: dict) -> None:
logger.warning("providers.model_catalog_cache_write_failed", error = str(exc))
+def cached_catalog_model_inputs(namespace: str, model: str) -> list[str] | None:
+ global _model_catalog_cache
+ if _model_catalog_cache is None:
+ _model_catalog_cache = _read_model_catalog_file()
+ bucket = ((_model_catalog_cache or {}).get("providers") or {}).get(namespace) or {}
+ entry = bucket.get(model.strip().lower())
+ inputs = entry.get("input") if isinstance(entry, dict) else None
+ return inputs if isinstance(inputs, list) else None
+
+
async def _fetch_models_dev_catalog() -> dict:
from core.inference.external_provider import _client
diff --git a/studio/backend/tests/test_qwencloud_provider.py b/studio/backend/tests/test_qwencloud_provider.py
new file mode 100644
index 00000000000..9e39e8599db
--- /dev/null
+++ b/studio/backend/tests/test_qwencloud_provider.py
@@ -0,0 +1,160 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+from __future__ import annotations
+
+import json
+import sys
+from pathlib import Path
+
+import pytest
+
+from core.inference.external_provider import ExternalProviderClient
+from core.inference.providers import get_provider_info, list_available_providers
+from models.providers import ProviderModelsRequest
+from routes import providers as providers_route
+from routes.inference import _external_takes_mcp_images
+
+sys.path.insert(0, str(Path(__file__).resolve().parent))
+from test_external_provider_sampling_over_the_wire import _Handler, _run, _Server # noqa: E402
+from test_external_provider_stream_usage_opt_in import _USAGE, _proxy, _usage_chunks # noqa: E402
+
+
+class _AuthHandler(_Handler):
+ def do_POST(self) -> None: # noqa: N802
+ self.server.auth = self.headers.get("Authorization") # type: ignore[attr-defined]
+ super().do_POST()
+
+
+_CHAT_IDS = [
+ "qwen3.8-max",
+ "qwen3.7-plus",
+ "qwen3.8-flash",
+ "qwen3.8-2.4t-a95b",
+ "qwen3.8-omni-flash",
+ "qwen3.5-omni-plus",
+ "qwen3-vl-plus",
+ "qwen3-coder-flash",
+ "qwen-mt-plus",
+ "qwen-plus-character",
+ "qwq-plus-2025-03-05",
+ "glm-5.3",
+ "ZHIPU/GLM-5.3",
+ "kimi-k3",
+ "deepseek-v4-pro",
+]
+_NON_CHAT_IDS = [
+ "pre-mock_new_gateway_session_affinity",
+ "qwen3.8-omni-flash-realtime",
+ "qwen3.5-livetranslate-flash-realtime",
+ "qwen3-livetranslate-flash",
+ "qwen-audio-3.1-realtime-plus",
+ "qwen-audio-3.0-asr-flash",
+ "qwen3-asr-flash-realtime",
+ "qwen3-tts-flash",
+ "qwen3-s2s-flash-realtime",
+ "qwen-image-3.0-pro",
+ "wan2.7-image-pro",
+ "z-image-turbo",
+ "qwen3.7-text-embedding",
+ "text-embedding-v4",
+ "qwen-vl-ocr-2025-11-20",
+ "tongyi-tingwu-slp",
+ "ccai-pro",
+]
+
+
+class _ModelsHandler(_Handler):
+ def do_GET(self) -> None: # noqa: N802
+ ids = self.server.ids # type: ignore[attr-defined]
+ body = json.dumps({"object": "list", "data": [{"id": i} for i in ids]}).encode()
+ self.send_response(200)
+ self.send_header("Content-Type", "application/json")
+ self.send_header("Content-Length", str(len(body)))
+ self.end_headers()
+ self.wfile.write(body)
+
+
+def test_the_qwencloud_model_list_keeps_only_chat_models(monkeypatch):
+ monkeypatch.setattr(
+ providers_route, "resolve_provider_api_key_or_400", lambda *a, **k: "sk-ws-test"
+ )
+ listed: list[str] = []
+ with _Server(handler = _ModelsHandler) as server:
+ server._httpd.ids = _NON_CHAT_IDS + _CHAT_IDS
+
+ async def go() -> None:
+ payload = ProviderModelsRequest(provider_type = "qwencloud", base_url = server.base_url)
+ result = await providers_route.list_provider_models(payload, "tester", False)
+ listed.extend(m.id for m in result)
+
+ _run(go)
+ assert listed == _CHAT_IDS
+
+
+def test_qwencloud_is_offered_with_its_own_endpoint():
+ rows = {row["provider_type"]: row for row in list_available_providers()}
+ qwencloud = rows["qwencloud"]
+ assert qwencloud["display_name"] == "QwenCloud"
+ assert qwencloud["base_url"] == "https://maas.qwencloudapi.com/compatible-mode/v1"
+ assert qwencloud["hidden"] is False
+ assert qwencloud["supports_studio_tools"] is True
+ assert {"qwen3.8-max", "qwen3.7-plus", "qwen3.8-flash"} <= set(qwencloud["default_models"])
+ assert rows["qwen"]["base_url"] == "https://dashscope-intl.aliyuncs.com/compatible-mode/v1"
+
+
+def test_a_qwencloud_connection_receives_the_usage_chunk():
+ chunks, body = _proxy("qwencloud", stream_options = {"include_usage": True})
+ assert body["stream_options"] == {"include_usage": True}
+ assert _usage_chunks(chunks) == [_USAGE]
+
+
[email protected]("enabled", [True, False])
+def test_the_qwencloud_thinking_toggle_reaches_the_wire(enabled):
+ with _Server(handler = _AuthHandler) as server:
+ client = ExternalProviderClient(
+ provider_type = "qwencloud", base_url = server.base_url, api_key = "sk-ws-test"
+ )
+
+ async def go() -> None:
+ async for _ in client.stream_chat_completion(
+ messages = [{"role": "user", "content": "hi"}],
+ model = "qwen3.7-plus",
+ enable_thinking = enabled,
+ ):
+ pass
+
+ _run(go)
+ body = server.bodies[-1]
+ assert server._httpd.auth == "Bearer sk-ws-test"
+ assert body["enable_thinking"] is enabled
+ assert "chat_template_kwargs" not in body
+ assert "thinking" not in body
+
+
+def test_qwencloud_vision_models_keep_mcp_pictures():
+ info = get_provider_info("qwencloud")
+ for model in info["default_models"]:
+ assert _external_takes_mcp_images("qwencloud", True, model, info) is True, model
+ assert _external_takes_mcp_images("qwencloud", True, "glm-5.3", info) is False
+
+
+def test_qwencloud_reads_vision_from_the_model_catalog(monkeypatch):
+ info = get_provider_info("qwencloud")
+ catalog = {
+ "fetched_at": 0,
+ "providers": {
+ "qwen": {
+ "qwen3-vl-plus": {"input": ["text", "image"]},
+ "qwen3.7-max": {"input": ["text"]},
+ }
+ },
+ }
+ monkeypatch.setattr(providers_route, "_model_catalog_cache", catalog)
+ assert _external_takes_mcp_images("qwencloud", True, "qwen3-vl-plus", info) is True
+ assert _external_takes_mcp_images("qwencloud", True, "qwen3.7-max", info) is False
+ assert _external_takes_mcp_images("qwencloud", True, "glm-5.3", info) is False
+ assert (
+ _external_takes_mcp_images("qwen", True, "qwen3-vl-plus", get_provider_info("qwen"))
+ is False
+ )
diff --git a/studio/frontend/scripts/check-bundle-budget.ts b/studio/frontend/scripts/check-bundle-budget.ts
index fc168f7d40f..7bc4734b1f2 100644
--- a/studio/frontend/scripts/check-bundle-budget.ts
+++ b/studio/frontend/scripts/check-bundle-budget.ts
@@ -200,8 +200,12 @@ export const BUDGET = {
// (6,800,980 -> 6,805,412 B raw, 2,045,645 -> 2,047,341 B transfer against 15e4bc8fde, one machine and build,
// Node 24, chunk count unchanged at 95): the shared slide module and SegmentedPill, imported by toggles that
// already live in startup chunks, so there is no separate module to lazy-load. Raised by that growth.
- transferBytes: 2_047_477,
- rawBytes: 6_805_847,
+ // QwenCloud as a connection (#14033) adds 167 B raw / 43 B transfer (6,804,234 -> 6,804,401 B raw, 2,047,526 ->
+ // 2,047,569 B transfer against ca3d1d0ce4, one machine and build, Node 24, chunk count unchanged at 95): its
+ // connection row, logo mapping and capability entries live in tables the startup chunks already load. Raised
+ // by that growth.
+ transferBytes: 2_047_520,
+ rawBytes: 6_806_014,
};
// The chunk count is reported but not budgeted. Splitting a page out of the entry raises it while lowering the
diff --git a/studio/frontend/src/features/chat/external-providers.ts b/studio/frontend/src/features/chat/external-providers.ts
index 2018a3ca773..c30a937bb68 100644
--- a/studio/frontend/src/features/chat/external-providers.ts
+++ b/studio/frontend/src/features/chat/external-providers.ts
@@ -282,7 +282,9 @@ export function providerModelSupportsVision(
// Mirrors _MIXED_CATALOG_PROVIDER_TYPES in studio/backend/routes/inference.py: vision for
// the family, unknown for a given model, so an MCP picture is not sent to one the
// registry says nothing about.
-const MIXED_CATALOG_PROVIDER_TYPES = new Set(["huggingface", "openrouter", "qwen"]);
+const MIXED_CATALOG_PROVIDER_TYPES = new Set(["huggingface", "openrouter", "qwen", "qwencloud"]);
+// Mirrors _CATALOG_VISION_NAMESPACES: mixed catalogs whose model catalog answers vision per model.
+const CATALOG_VISION_PROVIDER_TYPES = new Set(["qwencloud"]);
/** Whether the backend's external tool loop will hand this model an MCP picture -- the
* same rule as its _external_takes_mcp_images, so envelopes the backend would strip
@@ -298,7 +300,10 @@ export function providerModelTakesMcpImages(
const capability = REGISTRY_MODEL_CAPABILITIES.get(providerType)?.[modelId];
if (typeof capability?.vision === "boolean") return capability.vision;
}
- if (providerType && MIXED_CATALOG_PROVIDER_TYPES.has(providerType)) return false;
+ if (providerType && MIXED_CATALOG_PROVIDER_TYPES.has(providerType)) {
+ if (!CATALOG_VISION_PROVIDER_TYPES.has(providerType)) return false;
+ return modelCatalogSupportsVision(providerType, modelId) === true;
+ }
return true;
}
diff --git a/studio/frontend/src/features/chat/model-catalog.ts b/studio/frontend/src/features/chat/model-catalog.ts
index 53468936027..d692ddfe582 100644
--- a/studio/frontend/src/features/chat/model-catalog.ts
+++ b/studio/frontend/src/features/chat/model-catalog.ts
@@ -177,6 +177,7 @@ let mergedNamespaces: Map<string, Readonly<Record<string, ModelCatalogSnapshotEn
* way; without this every Codex lookup resolved to null and read as "not published". */
const CATALOG_NAMESPACE_ALIASES: Record<string, string> = {
openai_codex: "openai",
+ qwencloud: "qwen",
};
function catalogNamespace(providerType: string): string {
diff --git a/studio/frontend/src/features/chat/provider-capabilities.ts b/studio/frontend/src/features/chat/provider-capabilities.ts
index 42ebcb30601..c3c6f3e07d3 100644
--- a/studio/frontend/src/features/chat/provider-capabilities.ts
+++ b/studio/frontend/src/features/chat/provider-capabilities.ts
@@ -1258,6 +1258,7 @@ const CATALOG_REASONING_WIRE: Record<string, ReasoningWire> = {
kimi: { levels: [], off: "unified" },
deepseek: { levels: ["low", "high", "max"], aliases: { minimal: "low", medium: "high", xhigh: "high" }, off: "unified" },
qwen: { levels: [], off: "unified" },
+ qwencloud: { levels: [], off: "unified" },
huggingface: { levels: null, off: "unified" },
ollama: { levels: null, off: "unified" },
// A vLLM older than 0.21 400s on the wider levels; the backend then leaves them to the template default.
@@ -1432,6 +1433,7 @@ export function getExternalReasoningCapabilities(
case "ollama":
case "deepseek":
case "qwen":
+ case "qwencloud":
case "huggingface":
return catalogCapabilities(normalizedProvider, normalizedModel) ?? withEnableThinkingStyle();
case "vllm":
diff --git a/studio/frontend/src/features/chat/provider-logo-path.ts b/studio/frontend/src/features/chat/provider-logo-path.ts
index 7bc5c575956..f8a9a3b4fc6 100644
--- a/studio/frontend/src/features/chat/provider-logo-path.ts
+++ b/studio/frontend/src/features/chat/provider-logo-path.ts
@@ -12,6 +12,7 @@ const HUB_PROVIDER_IDS = new Map([
["huggingface", "huggingface"],
["kimi", "moonshotai"],
["qwen", "qwen"],
+ ["qwencloud", "qwen"],
]);
const HUB_LOGO_PATHS = new Map(
diff --git a/studio/frontend/tests/mcp-images.test.ts b/studio/frontend/tests/mcp-images.test.ts
index e9290d67d11..8bf5381a321 100644
--- a/studio/frontend/tests/mcp-images.test.ts
+++ b/studio/frontend/tests/mcp-images.test.ts
@@ -437,6 +437,10 @@ test("the upload gate is the backend's external MCP gate, not provider-level vis
assert.equal(providerModelTakesMcpImages("openrouter", "some/unknown-model"), false);
assert.equal(providerModelTakesMcpImages("huggingface", "org/unknown"), false);
assert.equal(providerModelTakesMcpImages("qwen", "qwen2.5-72b-instruct"), false);
+ assert.equal(providerModelTakesMcpImages("qwen", "qwen3-vl-plus"), false);
+ assert.equal(providerModelTakesMcpImages("qwencloud", "qwen3-vl-plus"), true);
+ assert.equal(providerModelTakesMcpImages("qwencloud", "qwen3.7-max"), false);
+ assert.equal(providerModelTakesMcpImages("qwencloud", "glm-5.3"), false);
// Known text-only endpoints stay a no; a vision family with a plain catalog is a yes.
assert.equal(providerModelTakesMcpImages("mistral", "mistral-large"), false);
assert.equal(providerModelTakesMcpImages("anthropic", "claude-x"), true);
diff --git a/studio/frontend/tests/model-catalog.test.ts b/studio/frontend/tests/model-catalog.test.ts
index ff359beb3ec..1bbe8f21146 100644
--- a/studio/frontend/tests/model-catalog.test.ts
+++ b/studio/frontend/tests/model-catalog.test.ts
@@ -177,6 +177,16 @@ test("Qwen only has a wire toggle, so a ladder collapses to on/off", () => {
assert.equal(getExternalReasoningCapabilities("kimi", "kimi-k2.6").supportsReasoning, true);
});
+test("QwenCloud reads the Qwen catalog for its thinking toggle and vision", () => {
+ for (const model of ["qwen3.8-max", "qwen3.7-plus", "qwen3.8-flash"]) {
+ const caps = getExternalReasoningCapabilities("qwencloud", model);
+ assert.equal(caps.reasoningStyle, "enable_thinking", model);
+ assert.equal(caps.supportsReasoning, true, model);
+ assert.equal(caps.supportsReasoningOff, true, model);
+ assert.equal(providerModelSupportsVision("qwencloud", model), true, model);
+ }
+});
+
test("Kimi K3's documented always-on effort ladder outranks the catalog's toggle", () => {
// models.dev lists a toggle, but the Kimi API has no thinking switch for kimi-k3.
const caps = getExternalReasoningCapabilities("kimi", "kimi-k3");
diff --git a/studio/frontend/tests/provider-logo-path.test.ts b/studio/frontend/tests/provider-logo-path.test.ts
index c889bf11eb9..397b2be4af5 100644
--- a/studio/frontend/tests/provider-logo-path.test.ts
+++ b/studio/frontend/tests/provider-logo-path.test.ts
@@ -19,6 +19,7 @@ test("connections use the model hub's canonical provider images", () => {
["huggingface", "huggingface"],
["kimi", "moonshotai"],
["qwen", "qwen"],
+ ["qwencloud", "qwen"],
]) {
const hubLogo = PROVIDER_LOGOS.find(({ id }) => id === hubId);
assert.ok(hubLogo, hubId);
@@ -37,7 +38,7 @@ test("Gemini keeps its own logo, separate from Google's hub image", () => {
test("every supported connection resolves to an existing public asset", () => {
for (const type of [
"openai", "openai_codex", "deepseek", "mistral", "huggingface", "kimi",
- "qwen", "gemini", "anthropic", "openrouter", "vllm", "ollama", "llama_cpp",
+ "qwen", "qwencloud", "gemini", "anthropic", "openrouter", "vllm", "ollama", "llama_cpp",
"liquid", "typesafe",
]) {
const path = providerLogoPath(type);