Qwen3.8-Flash

qwen3.8-flash

Qwen3.8-Flash is Qwen’s latest multimodal large model, combining strong understanding and generation capabilities with excellent response speed. The model natively supports a million-level context window and can process ultra-long documents, code repositories, and complex conversations in one go. It performs particularly well in coding assistance, agent collaboration, image-text understanding, and other scenarios—whether automatically fixing code, operating desktop applications, or analyzing charts and long videos, it can deliver accurate, high-quality results.

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
0.12/M	0.40/M	0.01/M	0.19/M

Overview

Input

Text Image

Output

Text

Qwen3.8-Flash API: Training cost only one-ninth of Qwen3.7-Plus, outperforming it on coding and office tasks

Qwen3.8-Flash is a lightweight multimodal model released by Alibaba's Tongyi Qianwen team on August 26, 2026. The open-source Qwen3.8-Flash-Next released the same day is its open-weight preview version, officially called an early preview of the Qwen4 architecture: 125B total parameters, only about 6B activated per token, using a hybrid structure that alternates Gated DeltaNet linear attention with the new Qwen Sparse Attention (QSA) at a 3:1 ratio. The hosted version qwen3.8-flash adds production features on this basis—a default 1 million token context and official built-in tools.

According to official statements, its training cost is about one-ninth that of Qwen3.7-Plus, yet it comprehensively surpasses the latter on coding and office tasks: SWE-bench Pro 62.5 (Qwen3.7-Plus 55.8), CoWorkBench 73.9 (65.1), JobBench 55.7 (27.6), Toolathlon Verified 73.5 (50.6). On the vision side, OSWorld 2.0 partial score improved from 21.5 to 52.3, and AndroidWorld reached 84.5. Compared with the previous generation Qwen3.6-Flash, maximum output increased from 65.5K to 128K tokens, and video was added as an input modality.

Things to know in advance: thinking mode is enabled by default, and the chain of thought can be up to 262,144 tokens long; multi-turn conversations retain historical thinking content by default (preserve_thinking). In thinking mode, the maximum input is 983,616 tokens, slightly lower than 991,808 tokens in non-thinking mode.

SeaWhale AI provides Qwen3.8-Flash through an OpenAI-compatible interface, supporting text, image, and video input, tool calling, structured output, and context caching.

Get API key · Model ID: qwen3.8-flash

Why choose Qwen3.8-Flash

Flash tier beats the previous generation Plus — SWE-bench Pro 62.5 vs. 55.8 for Qwen3.7-Plus, JobBench 55.7 vs. 27.6
Sparse MoE with 6B activated parameters — 125B total parameters, only 10+1 of 512 experts activated per token, inference cost at lightweight-model rates
1 million token context + 128K output — output limit is about 2x that of Qwen3.6-Flash (65.5K)
Native video input — unified text, image, and video input, LVBench 76.6
Major leap in GUI agent capabilities — OSWorld 2.0 partial score 52.3, 2.4x that of Qwen3.7-Plus (21.5)
Forerunner of the Qwen4 architecture — QSA sparse attention only lets 1/4 of layers store growing KV, saving more memory in long contexts

Core capabilities

01 Agentic coding

Qwen3.8-Flash is one of the models with the highest coding scores in the Qwen3.8 series. Officially measured under the Claude Code framework, SWE-bench Pro is 62.5, higher than DeepSeek-V4-Flash's 56.0 and Claude Opus 4.6's 53.4; SWE-bench Multilingual 81.0, LiveCodeBench v6 91.9.

Repository-level bug fixing: DeepSWE 1.1 reaches 58.7, while Qwen3.7-Plus is only 16.5
Multilingual engineering: SWE-bench Multilingual 81.0, 5.2 points higher than Qwen3.7-Plus
Algorithms and competitive programming: LiveCodeBench v6 91.9

02 Office and tool orchestration

At launch, the official team listed "office tasks" as a main focus alongside coding. CoWorkBench, aimed at real office workflows, scores 73.9, and JobBench, aimed at occupational tasks, scores 55.7, the latter being 2x that of Qwen3.7-Plus.

Multi-tool collaboration: Toolathlon Verified Pass@1 73.5
Instruction following: IFBench 81.3, higher than Claude Opus 4.6's 62.5
Supports Function Calling, structured output, and prefix continuation; web search is supported in the Beijing and Singapore regions

03 Screen understanding and GUI operation

Building on Qwen3.7-Plus's visual agent, Flash pushes GUI operation capabilities up another notch. OSWorld 2.0 binary/partial scores are 19.4 / 52.3, while Qwen3.7-Plus has only 2.8 / 21.5; AndroidWorld 84.5, ClawEval-MM Pass@3 64.4.

Desktop and mobile interface automation
Design mockup to webpage: Vision2Web 64.0 (Qwen3.7-Plus is 42.1)
Real-world scene understanding: RealWorldQA 88.5, ERQA 72.3

04 Long context and video understanding

The hosted version defaults to a 1 million token context, which can hold a large code repository or extracted frames from a long video in one go. QSA splits the sequence into micro-blocks and applies attention only to the most relevant 2,048 token positions; of the 48 layers, only 12 need to store KV that grows with length, while the remaining 36 layers are fixed-size states.

Long video understanding: LVBench 76.6
Chart reasoning: CharXiv RQ (using code interpreter) 90.6, MathVision (using code interpreter) 95.7
Supports context caching, making repeated calls with long prefixes more economical

Best use cases

Use case	Description
Coding agent	Low-cost execution of repository-level modification tasks like Claude Code / Cline
Office automation	Workflows chaining multiple tools such as documents, spreadsheets, and email
GUI automation	Read the screen and plan click paths, on desktop and Android
Video and image analysis	Long video summarization, chart interpretation, on-site image judgment
Long document / large repository Q&A	1 million token context holds all materials at once
High-concurrency production traffic	6B activated parameters, suitable for batch processing and online services requiring large numbers of calls

Differences between Qwen3.8-Flash, Qwen3.6-Flash, and Qwen3.7-Plus

Capability	Qwen3.8-Flash	Qwen3.6-Flash	Qwen3.7-Plus
Model ID	`qwen3.8-flash`	`qwen3.6-flash`	`qwen3.7-plus`
Release date	August 26, 2026	—	—
Architecture	Sparse MoE (125B total / 6B activated) + QSA	Linear attention + sparse MoE	Dense
Context window	1 million tokens	1 million tokens	1 million tokens
Maximum output	128K tokens	65.5K tokens	128K tokens
Input modalities	Text, image, video	Text, image	Text, image
SWE-bench Pro	62.5	—	55.8
CoWorkBench	73.9	—	65.1
OSWorld 2.0 (partial score)	52.3	—	21.5
GPQA Diamond	91.7	—	90.3
Positioning	Lightweight tier, mainstay for coding and office work	Speed-first lightweight model	Cost-effective multimodal agent

Benchmark data comes from the official Qwen model card (Qwen3.8-Flash-Next open-weight version) and reflects vendor self-reported results; the hosted version and open-weight version are not exactly the same product. Specific billing is subject to the real-time price card at the top of the page.

FAQ

When was Qwen3.8-Flash released? August 26, 2026. On the same day, Alibaba also open-sourced Qwen3.8-Flash-Next (Hugging Face, ModelScope, qwen-community-1.0 license), and the hosted version qwen3.8-flash subsequently launched on Alibaba Cloud Model Studio and Qwen Cloud.

What has been upgraded compared with Qwen3.6-Flash and Qwen3.7-Plus? The architecture has changed to the Qwen4 preview architecture (Gated DeltaNet + QSA sparse attention + sparse MoE), maximum output increased from 65.5K to 128K tokens, and video input was added. In terms of capabilities, it surpasses the higher-tier Qwen3.7-Plus on most official benchmarks such as SWE-bench Pro, CoWorkBench, JobBench, and OSWorld 2.0, while the official team claims training cost is only one-ninth of the latter.

What are the context and output limits? Context 1 million tokens; maximum output 128K tokens (131,072), the same in thinking and non-thinking modes. Thinking mode additionally has a chain of thought of up to 262,144 tokens; in this case, the maximum input is 983,616 tokens.

How is thinking mode configured? Enabled by default. Use enable_thinking to toggle thinking; preserve_thinking controls whether multi-turn conversations retain previous thinking content (retained by default). The open-weight model card also provides reasoning_effort (xhigh by default; medium and low are optional), but the official team notes: in multi-turn agent tasks, lowering reasoning effort does not necessarily shorten overall completion time.

What tools and features are supported? Supports Function Calling, structured output, context caching, and prefix continuation; web search is available only in the Beijing and Singapore regions. Input supports text, image, and video; output is text.

Should I choose Qwen3.8-Flash or Qwen3.7-Plus? According to official benchmarks, Qwen3.8-Flash leads overall in coding, office work, and GUI operation, and belongs to the cheaper Flash tier, so new projects should prioritize it. On a few individual metrics, Qwen3.7-Plus still has a slight edge, for example CharXiv RQ without using a code interpreter (85.8 vs. 84.6).

Where does it lag behind? Yes. HLE 35.9, lower than Claude Opus 4.6's 40.0; NL2Repo-Bench for generating repositories from scratch is 48.1, lower than DeepSeek-V4-Flash's 54.2. In addition, almost all public benchmarks are Qwen's own tests, and independent third-party evaluations are currently lacking; its architecture is a preview of Qwen4, and later versions may change.

Why use the Qwen3.8-Flash API with SeaWhale AI

No Alibaba Cloud account required — call directly with one key, with no need to build your own 125B-parameter inference cluster
OpenAI-compatible interface — integrate by changing the model name in existing code; video and image input formats are consistent with OpenAI
Switch within the same series as needed — share a key with Qwen3.7-Max and Qwen3.7-Plus, and route by task difficulty
RMB billing — domestic models are priced in RMB

API

API integration

Model IDUse this value as the model in inference requests

qwen3.8-flash

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-flash 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-flash",
    "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-flash",
    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-flash',
  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 ?? '')
}