Available Models List

CloudBase AI comes with built-in large language models from major providers. You can call them directly through the Mini Program, Web SDK, Node SDK, HTTP API, and more (see Invocation Methods) — there is no need to integrate each provider yourself.

The built-in models' capabilities are aligned with each vendor's official models: CloudBase standardizes the invocation path, multi-platform SDKs, and protocols, but does not trim, replace, or alter the models' own capabilities. The capabilities, parameters, and limits described in each vendor's official documentation therefore apply on CloudBase as well.

The models actually available and their prices are subject to what is shown in the CloudBase console. For models that are no longer served, see Deprecated Models List.

How to read the tables

• ✅ means the vendor's official documentation explicitly states support for that capability; ❌ means the official documentation explicitly states it is not supported; — means the vendor does not state it for that specific model, which does not necessarily mean it is unsupported.
• "Context" is the model's maximum context window, provided as a selection reference.
• Each vendor section links to its official documentation, which is the source of the capability values. If a value differs from the vendor's latest documentation or the CloudBase console, the latter prevails.

Models and Capabilities

Hunyuan

Official documentation: Hunyuan invocation guide

Model	Context	Image	Video	File	Deep Thinking	Tool Calling	Structured Output	Caching
hy4-preview	1M	❌	❌	—	✅	✅	✅	✅
hy3	256K	❌	❌	—	✅	✅	✅	✅
hy-vision-2.0-instruct	44K	✅	❌	—	—	—	—	—
hunyuan-t1-vision-20250916	40K	✅	❌	—	✅	—	—	—
hunyuan-turbos-vision-video-20250728	32K	❌	✅	—	—	—	—	—
hunyuan-role-latest	32K	—	—	—	—	—	—	—
hy-role	32K	—	—	—	—	—	—	—
hy-mt2-lite	8K	—	—	—	—	—	—	—
hy-mt2-plus	8K	—	—	—	—	—	—	—
hy-mt2-pro	8K	—	—	—	—	—	—	—

Of these, hunyuan-role-latest and hy-role target role-playing, while hy-mt2-lite, hy-mt2-plus, and hy-mt2-pro target translation. They are tuned for specific tasks, so the capability columns above do not apply.

• Model specs: The TokenHub model list gives each model's capability support, context window, and max input / output
• Deep thinking: hy4-preview and hy3 have deep thinking on by default and support Preserved Thinking; toggle it with thinking.type, and set reasoning_effort to low / high (default high). With tools, hy3 adapts to task complexity and maps low to high
• Structured output: Emit JSON per a JSON Schema via response_format; see the Hunyuan invocation guide
• Image and video: Multimodal understanding · Video understanding; the Hunyuan language models in the table above do not accept image / video input, so use these vision models for multimodal scenarios

Hunyuan vision models

• hy-vision-2.0-instruct: A fast-thinking image-to-text model for general image-to-text scenarios. Compared with the previous generation, it delivers notable improvements in basic perception, content recognition, knowledge, OCR, STEM, reasoning, and chart understanding.
• hunyuan-t1-vision-20250916: A deep-thinking image-to-text model for general image-and-text Q&A, visual grounding, OCR, charts, question solving from photos, and image-based creation. English and minor-language capabilities are also improved.
• hunyuan-turbos-vision-video-20250728: A video understanding model that supports basic video understanding capabilities such as video description and video content Q&A.

DeepSeek

Official documentation: DeepSeek API docs

Model	Context	Image	Video	File	Deep Thinking	Tool Calling	Structured Output	Caching
deepseek-v4.1-flash	1M	✅	—	—	✅	✅	✅	✅
deepseek-v4.1-flash-1017	1M	✅	—	—	✅	✅	✅	✅
deepseek/deepseek-flash	1M	✅	—	—	✅	✅	✅	✅
deepseek-v4-pro	1M	❌	—	—	✅	✅	✅	✅
deepseek-v4-pro-0813	1M	❌	—	—	✅	✅	✅	✅

Of these, deepseek/deepseek-flash corresponds to flash-v4.1 and deepseek-v4.1-flash-1017 is the 1017 snapshot of that model; deepseek-v4-pro is the preview version and deepseek-v4-pro-0813 is the general-availability version.

• Image understanding: Image Understanding
• Deep thinking: Thinking and non-thinking modes can be switched; see Thinking Mode
• Structured output: JSON Output

Zhipu AI

Official documentation: Zhipu AI documentation

Model	Context	Image	Video	File	Deep Thinking	Tool Calling	Structured Output	Caching
glm-5.3-flashx	1M	✅	✅	✅	✅	✅	✅	✅
glm-5.3-flash	1M	✅	✅	✅	✅	✅	✅	✅
glm-5.3	1M	—	—	—	✅	✅	✅	✅
glm-5.2	1M	—	—	—	✅	✅	✅	✅

• Vision and files: GLM-5.3-Flash / FlashX; only these two models currently accept document file input, and only via URL
• Deep thinking: thinking.type only accepts enabled; reasoning_effort accepts low / high / max
• Structured output: Structured Output

Moonshot AI

Official documentation: Kimi Chat Completions API

Model	Context	Image	Video	File	Deep Thinking	Tool Calling	Structured Output	Caching
kimi-k3	1M	✅	✅	—	✅	✅	✅	✅
kimi-k2.8-preview	1M	✅	✅	—	✅	✅	✅	✅
kimi-k2.7-code-highspeed	256K	✅	✅	—	✅	✅	✅	✅
kimi-k2.7-code	256K	✅	✅	—	✅	✅	✅	✅
kimi-k2.6	256K	✅	✅	—	✅	✅	✅	✅

• Vision: Kimi vision models
• Deep thinking: kimi-k3 always reasons; top-level reasoning_effort accepts low / high / max (default max). K2.x uses thinking.type (enabled / disabled); kimi-k2.7-code cannot disable thinking
• Structured output: response_format; see Chat Completions API

MiniMax

Official documentation: Model invocation

Model	Context	Image	Video	File	Deep Thinking	Tool Calling	Structured Output	Caching
minimax-m3.1-flash	1M	✅	✅	—	✅	✅	—	✅
minimax-m3	1M	✅	✅	—	✅	✅	—	✅

• Vision: OpenAI-compatible API
• Deep thinking: thinking cannot be disabled on MiniMax-M3.1-Flash; adjust depth with effort (low / medium / high / xhigh / max, default max). MiniMax-M3 can disable thinking via thinking.type; M2.x cannot

Xiaomi MiMo

Official documentation: MiMo open platform docs

Model	Context	Image	Video	File	Deep Thinking	Tool Calling	Structured Output	Caching
mimo-v2.6-pro	1M	✅	✅	—	✅	✅	✅	✅
mimo-v2.6-flash	1M	✅	✅	—	✅	✅	✅	✅

• Model details: MiMo-V2.6-Pro, covering multimodal understanding, deep thinking, and structured output parameters

StepFun

Official documentation: StepFun open platform docs

Model	Context	Image	Video	File	Deep Thinking	Tool Calling	Structured Output	Caching
step-5-preview	1M	✅	✅	—	✅	✅	✅	✅

• Model details: Step 5 Preview, covering multimodal understanding and JSON Mode / JSON Schema
• Deep thinking: reasoning_effort accepts low / medium / high

Invocation Methods

CloudBase offers several ways to call the models, covering Mini Program, Web, server-side, and third-party SDKs. Pick the one that fits your application; detailed usage of each capability lives on the corresponding page under "Feature Guide".

Method	Description
Mini Program	WeChat Mini Program, via wx.cloud.extend.AI; no SDK installation required
wx-server-sdk	WeChat Cloud Functions, via the built-in wx-server-sdk
Web SDK	Browser / web apps, via @cloudbase/js-sdk
Node SDK	Node.js server side (Cloud Functions, Cloud Run), via @cloudbase/node-sdk
cURL	Direct HTTP API access for backend services, scripts, and any programming language
OpenAI SDK	Reuse your existing OpenAI SDK for easy migration and multi-model switching
Anthropic SDK	Reuse your existing Anthropic SDK — just replace authToken and baseURL

note

The Mini Program, wx-server-sdk, Web SDK, and Node SDK methods authenticate through your CloudBase login state and environment, while cURL, the OpenAI SDK, and the Anthropic SDK require an API Key first.

FAQ

The same capability takes different parameter names across models. How do I keep this consistent?

CloudBase uses the OpenAI Chat Completions-compatible protocol throughout (Responses API and Anthropic Messages API are also supported), so message structure, tools, image_url, and similar fields are identical on CloudBase. Vendor-specific parameters — such as Hunyuan's thinking levels or Zhipu's thinking.type — follow the relevant official documentation and can be passed through as needed.

Capabilities match the official service — does that mean this page replaces the vendor docs?

No. This page answers "model selection — which models exist and which one can do what"; the vendor documentation answers "implementation — which parameters to set and what the limits are". Use them together: pick the model here, then confirm the parameter details in the vendor documentation.

If a capability is marked —, does that mean it is unsupported?

Not necessarily. — only means the vendor does not state that capability for that specific model, which is common for models the vendor positions as text-only. To determine whether a capability is available, check the model's official documentation, or simply call it from the console to verify.