KovAPIKovAPI
AnnouncementsAll updates →
NoticeNezhaGate is now KovAPI — new address kovapi.com NoticeCompensation for recent video generation failures NewSeedance 2.0 Mini is back and 2.0 Fast adds 480P, at a launch price
简体中文English ✓Español한국어日本語Русский
Log in Get API Key
Chat ● Operational Service status →

Qwen3.8 Max

⧉
qwen3.8-max
Model Type MaxFlash

Qwen3.8 Max is the flagship of Alibaba Cloud's Qwen 3.8 generation and the strongest Qwen model for reasoning, coding and agent work, with dependable Chinese and English. It always thinks before it answers, with the reasoning in reasoning_content, and it supports tool calling, JSON output, streaming and image input (send images as base64 data URIs). Fully OpenAI-compatible: set model to qwen3.8-max. Pay-as-you-go, failed calls never billed.

ChatReasoningCodingImage inputOpenAI compatible
Get API Key → API Docs

Why this model

Alibaba's Qwen 3.8 flagship: always thinks, with tool calling, JSON output and image input

Highlights

  • The flagship of Alibaba Cloud's Qwen 3.8 generation, for reasoning, coding and agent work
  • Always thinks before it answers; the reasoning comes back in reasoning_content
  • Tool calling, JSON output and streaming over the OpenAI-compatible API
  • Text and image input
  • One family: the pinned 0902 snapshot and the low-cost, high-volume Qwen3.8 Flash

Price vs. official list

KovAPIList price Input$1.4$2 Output$4.2$6

Per 1M tokens, USD · You save ↓30%

Live Test · Playground

Try out Qwen3.8 Max right here (available after login).

Input

Advanced
Web search
Memory · multi-turn

Conversation

Start a conversationType a message below to begin

About Qwen3.8 Max

Qwen3.8 Max is the flagship of Alibaba Cloud's Qwen 3.8 generation, served on KovAPI through the OpenAI-compatible API. It is strong at reasoning, coding and agent work, with dependable Chinese and English and image input. It always thinks before it answers: the reasoning comes back in message.reasoning_content and the answer in content, so clients that only read content need no changes. Tool calling, JSON output and streaming are supported. Set model to qwen3.8-max; pay-as-you-go, failed calls never billed.

Use cases

Complex reasoning and maths

It thinks before it answers, which suits multi-step reasoning, maths and logic problems; the reasoning is visible in reasoning_content.

Code generation and review

Writing, refactoring, explaining and reviewing code, with response_format for dependable JSON output.

Agents and tool calling

OpenAI-format tools / tool_calls plug into any agent framework for retrieval, function calls and multi-step tasks.

Image understanding

Q&A over screenshots, receipts and charts, reading the text in an image accurately; send images as base64 data URIs.

How to choose

Choose Qwen3.8 Max when you need the strongest Qwen reasoning, coding and agent work; choose the 0902 snapshot for a pinned version and reproducible output; choose Qwen3.8 Flash for high volume when cost and speed matter most. All three belong to the Qwen3.8 family, so switching is just the model field.

FAQ

Can I turn thinking off or limit its length?
No. Through KovAPI, Qwen3.8 Max always thinks, and parameters such as enable_thinking, reasoning_effort or thinking_budget neither turn it off nor shorten it. Reasoning tokens are billed at the output rate inside usage.completion_tokens. For shorter replies, ask for concise answers in the prompt.
How do I send an image?
Put an image_url with a base64 data URI (data:image/png;base64,...) in the message content. In our tests Qwen3.8 Max read the text in an image accurately; public image links currently fail, so convert them to base64 first, or use Qwen3.8 Flash, which accepts links directly.
Is prompt caching supported?
We have not seen cache hits on Qwen3.8 Max so far, so its input bills at the full input rate. If you reuse the same long prefix, such as one long document, choose the 0902 snapshot: it cached automatically in both of our test runs, and the cached part settles at the cache rate. Qwen3.8 Flash accepts cache_control, but hits are not guaranteed.
What should I know about long answers?
The model thinks first, and even a simple question occasionally thinks for a long time (our longest was about 2 minutes). For long output or latency-sensitive calls use stream=true: the reasoning streams in delta.reasoning_content and the answer in delta.content.
Am I charged for failed calls?
No. Only calls that return normally are settled, at actual usage; upstream errors and timeouts never touch your balance.

Related Models

Explore other models you can integrate.

View all →
Qwen3.8 Flash Chat

Qwen3.8 Flash

qwen3.8-flash

The light, fast Qwen3.8 - high volume, low cost

$0.10/1M in · $0.33/1M out ↓31% View →
GLM-5.3 Chat

GLM-5.3

glm-5.3

The new Z.ai flagship - coding and agents

$1.12/1M in · $3.52/1M out ↓20% View →
DeepSeek V4.1 Flash Chat

DeepSeek V4.1 Flash

deepseek-v4.1-flash

DeepSeek's current Flash model - thinking on or off

$0.07/1M in · $0.30/1M out ↓50% View →