Qwen3.8-Max

qwen3.8-max

A 2.4-trillion-parameter MoE flagship with fully upgraded coding and office capabilities, capable of autonomously coding for over ten days to deliver a complete project. It handles hundreds of professional tasks across law, finance, design, and more, delivering production-grade results end-to-end in a single conversation. Native visual understanding runs through the entire planning, execution, and verification workflow, supporting deep semantic analysis of ultra-long documents and long videos. In long-horizon tasks, it autonomously plans and iterates in a closed loop, continuously evolving.

Context window1.0M

Provider通义千问

Released2026/10/20

Playground

Pricing

The same model is available through multiple service channels — choose based on latency, reliability and cost.

Prices in $ / 1M tokens

To pick a channel, add a provider field to the request body, for example "provider": { "channel": "direct" }. Valid values are direct / stable / economical; omit it to use the default channel.

Direct

Direct upstream connection — best when you need native behavior and the full context window.

Input	Output	Cache read	Cache write
1.79/M	5.36/M	0.15/M	2.23/M

Overview

Input

Text Image

Output

Text

Qwen3.8-Max API: The 2.4-Trillion-Parameter Multimodal Agent Flagship, with a Full Generational Leap in Coding and Office Work

Qwen3.8-Max is the flagship model in Alibaba's Tongyi Qianwen Qwen3.8 series. It was previewed at the World Artificial Intelligence Conference (WAIC) on July 19, 2026, and officially launched its API on August 3. It uses a sparse MoE architecture with 2.4 trillion total parameters and about 95 billion activated per token, natively supports text, image, and video input, has a 1-million-token context window, and is positioned as a multimodal agent model for coding, real-world office work, and long-horizon tasks.

Compared with the previous generation Qwen3.7-Max, the improvement is almost generational: in official evaluations, OSWorld-Verified rose from 73.3 to 86.1, TerminalBench-2.1 from 74.5 to 86.6, PaperBench from 64.8 to 93.0, and FrontierSWE from 40.7 to 73.5. Qwen3.7-Max only accepts text input, while Qwen3.8-Max brings visual capabilities into the execution process, enabling it to inspect screenshots of its own output during tasks and read PDFs of more than 200 pages. The 0902 post-training update released on September 2 raised all 8 coding benchmarks, and Code Arena WebDev ranked first with 1691 points.

What you need to know in advance: Qwen3.8-Max enables deep thinking by default (reasoning_effort defaults to xhigh), outputs tend to be long, and speed is relatively slow. For latency-sensitive scenarios, it is recommended to lower the level.

SeaWhale AI provides Qwen3.8-Max through an OpenAI-compatible API, supporting tool calling, structured output, image and video input, and reasoning_effort reasoning intensity control.

Get API Key · Model ID: qwen3.8-max

Why Choose Qwen3.8-Max

OSWorld-Verified 86.1 — 12.8 points higher than Qwen3.7-Max, surpassing Claude Opus 4.8's 83.4 in official tests
TerminalBench-2.1 86.6 — 12.1 points higher than the previous generation, with terminal agent capabilities entering the first tier
Code Arena WebDev No. 1 — 1691 points after the 0902 update, higher than Claude Opus 5 Max's 1687
Native multimodality — unified input of text, images, and video; the previous generation Max only supported text
1-million-token context + 128K output — large repositories and ultra-long documents fit in one pass
Low unit price — combined input and output pricing is less than one-third that of Claude Opus 5

Core Capabilities

01 Long-Horizon Agentic Coding

Qwen3.8-Max is designed for long-duration autonomous execution. In an official demo, it autonomously developed a CLI tool for about 16 consecutive days, made 265 commits, opened 127 PRs, handled 151 issues, and before each merge ran builds, unit tests, end-to-end tests, and CI. After the 0902 update, DeepSWE 1.1 rose from 56.6 to 69.3, NL2Repo-Bench from 55.9 to 64.9, and Terminal-Bench 3.0 from 11.3 to 29.0.

Repository-level feature implementation and issue fixing
Multi-day autonomous development based on issue / PR workflows
Front-end and web application generation (No. 1 on Code Arena WebDev)

02 Real-World Office and Research Tasks

On benchmarks measuring real-world work deliverables, Qwen3.8-Max shows the largest improvement over the previous generation: JobBench rose from 31.3 to 53.4 (0902 further to 64.0), and CoWorkBench from 64.6 to 74.8. In the official paper-reproduction demo, it reproduced 6 paper conclusions in about 125 hours and over 1,100 operations, and scored 2.71 points higher than the original paper's method on AIME24.

Research and analysis across documents and spreadsheets
Paper reproduction and experiment iteration (PaperBench 93.0)
Instruction following (IFBench 82.8), professional finance and medical tasks

03 Vision-Driven Agents

Visual capabilities are not only for looking at images and answering questions: Qwen3.8-Max takes screenshots to check its own output during execution and then makes corrections based on them. OSWorld-Verified 86.1, MobileWorld 77.8 (previous generation 51.2), and ERQA 77.8 all lead Claude Opus 4.8 in official comparisons. For long-video understanding, a video memory graph enables processing of more than 100 hours of footage, with LVBench 81.8.

Automated operation of computer and mobile interfaces
Understanding of 200+ page PDFs and charts (CharXiv RQ 88.4)
Long-video retrieval and analysis

04 Adjustable Reasoning Effort and Engineering Interfaces

reasoning_effort offers three levels: xhigh (default, complex tasks), medium (balanced accuracy and speed), and low (prioritize speed and cost). It also supports Function Calling, structured output, prefix continuation, and implicit and explicit context caching; the read price for explicit cache is far lower than the normal input price.

Switch among three reasoning effort levels by task
Tool calling and JSON structured output
Implicit / explicit caching reduces the cost of multi-turn agent calls

Best Use Cases

Scenario	Description
Agentic coding	Repository-level development and fixing; can connect to Claude Code, Codex, Qwen Code
Front-end and Web development	Ranked first on Code Arena WebDev
Computer / mobile automation	OSWorld-Verified 86.1, vision-driven interface operation
Research and office deliverables	Report writing, data analysis, paper reproduction
Long-document and long-video analysis	1-million-token window, 200+ page PDFs, 100+ hours of video
Domestic flagship selection	First tier in comprehensive capability among domestic models, RMB settlement

Differences Between Qwen3.8-Max, Qwen3.7-Max, and GLM-5.3

Capability	Qwen3.8-Max	Qwen3.7-Max	GLM-5.3
Model ID	`qwen3.8-max`	`qwen3.7-max`	`GLM-5.3`
Release date	August 3, 2026 (`0902` update on September 2)	May 20, 2026	August 14, 2026
Total parameters / active	2.4 trillion / about 95 billion	Trillion-level MoE	753 billion / about 40 billion
Input modalities	Text, images, video	Text	Text
Context window	1 million token	1 million token	1 million token
Max output	128K token	128K token	128K token
Terminal-Bench 3.0	29.0 (`0902`)	—	28.3
OSWorld-Verified	86.1	73.3	—
AA Intelligence Index	45.42 (`0902`)	—	44.78 (max)
Reasoning effort	low / medium / xhigh	—	low / high / max
Positioning	Multimodal agent flagship	Previous-generation text-only flagship	Coding and safety flagship

Comparison data for Qwen3.8-Max and Qwen3.7-Max comes from Qwen's official release; Terminal-Bench 3.0 comes from each vendor's official release; AA Intelligence Index is a third-party evaluation by Artificial Analysis (v4.3.2, October 2026). Specific billing is subject to the real-time pricing card at the top of the page.

FAQ

When was Qwen3.8-Max released? Previewed at WAIC on July 19, 2026, and officially launched its API on August 3. The 0902 post-training update was released on September 2; architecture, parameters, and context remain unchanged, while coding and agent capabilities are further improved.

What has been upgraded compared with Qwen3.7-Max? Three aspects: first, it changed from text-only to native support for image and video input; second, agent capabilities improved significantly, with OSWorld-Verified 73.3 → 86.1, TerminalBench-2.1 74.5 → 86.6, and FrontierSWE 40.7 → 73.5; third, real-world office deliverables are stronger, with JobBench 31.3 → 53.4 and PaperBench 64.8 → 93.0.

What are the context and output limits? The context window is 1 million token, and the maximum output is 128K token (131,072). When thinking is enabled, the maximum input is about 984,000 token, and the chain of thought can be up to 262,144 token.

How is reasoning effort set? Set via reasoning_effort; options are low, medium, and xhigh, with a default of xhigh. Keep the default for complex coding and research tasks; for latency-sensitive tasks such as chat, classification, and extraction, it is recommended to use low or medium, which can significantly reduce output token and wait time.

What tools and features are supported? Supports Function Calling, structured output, prefix continuation, implicit and explicit context caching; input supports text, images, and video. The API is OpenAI-compatible and can also be connected to coding tools such as Claude Code and Codex.

Where does Qwen3.8-Max lag behind? Yes. First, it is slow and produces long outputs: in Artificial Analysis measurements, median output speed is about 37 token/s, and running the full evaluation suite used about 190 million output tokens, making the per-task cost actually higher than Claude Opus 5 (xhigh). Second, hallucination control is relatively weak, with an AA-Omniscience score of 3.4, far below Claude Opus 5's 37.1. Third, it still lags on the hardest software engineering tasks: SWE-bench Pro 67.7 is lower than Claude Fable 5's 80.0, and Terminal-Bench 3.0 29.0 is lower than Claude Opus 5's 42.7. In addition, the official TerminalBench result used its own test framework and a 5-hour timeout; third-party retests under standard frameworks score lower.

Should I choose Qwen3.8-Max or GLM-5.3? If you need to look at images, watch videos, operate computer interfaces, or do long research and office tasks, choose Qwen3.8-Max; if you only do text-only coding or code security auditing and care more about fewer output token and faster responses, GLM-5.3 is more suitable. Their Terminal-Bench 3.0 scores are close (29.0 vs 28.3), and you can compare them directly under the same SeaWhale AI account using the same code.

Why Choose SeaWhale AI to Use the Qwen3.8-Max API

No Alibaba Cloud account activation required — call directly with one key, no need to separately activate Bailian
OpenAI-compatible API — existing code can be integrated with two lines changed
Compare side by side with other flagships — compare performance and cost with GLM, Kimi, and DeepSeek under the same account
RMB settlement — domestic models are priced in RMB, with clear accounting

API

API integration

Model IDUse this value as the model in inference requests

qwen3.8-max

API KeyBearer token used to authenticate inference requests

Manage API keys

Base URLOpenAI compatible · /chat/completions

https://api.haijingai.com/v2/

provider OptionalSelects a service channel; omit it and the system picks the default

"provider": { "channel": "direct" }

qwen3.8-max usage examples

SeaWhale AI is compatible with the OpenAI API protocol, so you can call it with the OpenAI SDK or plain HTTP requests. Streaming is enabled by default.

cURLPythonNode.js

curl https://api.haijingai.com/v2/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer <API_KEY>" \
  -d '{
    "model": "qwen3.8-max",
    "messages": [
      {"role": "system", "content": "You are a helpful assistant."},
      {"role": "user", "content": "Hello!"}
    ],
    "provider": { "channel": "direct" },
    "stream": true
  }'
# provider is optional — remove this line to use the default channel

from openai import OpenAI

client = OpenAI(
    base_url="https://api.haijingai.com/v2",
    api_key="<API_KEY>",
)

stream = client.chat.completions.create(
    model="qwen3.8-max",
    messages=[
        {"role": "system", "content": "You are a helpful assistant."},
        {"role": "user", "content": "Hello!"},
    ],
    stream=True,
    # Optional: pick a service channel; omit to use the default
    extra_body={"provider": {"channel": "direct"}},
)

for chunk in stream:
    if chunk.choices[0].delta.content:
        print(chunk.choices[0].delta.content, end="", flush=True)

import OpenAI from 'openai'

const client = new OpenAI({
  baseURL: 'https://api.haijingai.com/v2',
  apiKey: '<API_KEY>',
})

const stream = await client.chat.completions.create({
  model: 'qwen3.8-max',
  messages: [
    { role: 'system', content: 'You are a helpful assistant.' },
    { role: 'user', content: 'Hello!' },
  ],
  stream: true,
  // Optional: pick a service channel; omit to use the default
  // @ts-expect-error provider is a SeaWhale AI extension, not in the OpenAI SDK types
  provider: { channel: 'direct' },
})

for await (const chunk of stream) {
  process.stdout.write(chunk.choices[0]?.delta?.content ?? '')
}