MoeRouter Blog
English
Get an API key

Qwen3.8-Flash and Qwen3.8-Max review roundup: Alibaba's light tier and its flagship

Qwen3.8 has two models people use: the light Qwen3.8-Flash and the flagship Qwen3.8-Max (0902 build). Alibaba's scores, an independent Chinese benchmark and Bilibili creators' tests: what each is good at, where it falls short, the price.

qwen3.8-flashqwen3.8-max-0902

In short

  • Qwen3.8-Max was released on August 3, 2026 and upgraded to the 0902 build on September 2 at an unchanged list price of $2 in / $6 out per 1M tokens. Qwen3.8-Flash is the light multimodal model released in late August.
  • The 0902 build is first on Code Arena's front-end coding board with 1,691 points; in Alibaba's own comparison table it still trails Claude Opus 5 on 8 of 14 benchmarks.
  • Qwen3.8-Flash scores 74.0% on ReLE, an independent Chinese benchmark, just under GLM-5.3-Flash in its price class. In hands-on tests it passed the reasoning puzzles and left small interface bugs in coding projects.
  • Alibaba released the newer Qwen 4 in November. MoeRouter offers the Qwen3.8 line at present, where qwen3.8-flash is 9th by usage.
On this page
  1. What the two models are
  2. What they cost on MoeRouter
  3. Benchmarks
  4. Hands-on
  5. Which to pick
  6. How to call them
  7. Frequently asked questions

What the two models are#

Qwen3.8-Max is Alibaba's flagship, released on August 3, 2026: a 2.4-trillion-parameter mixture-of-experts model with a 1M-token context and text, image and video input, aimed at autonomous software engineering and long-horizon enterprise work[3]. On September 2 it was upgraded to the 0902 build: the same base model with further post-training for coding and professional office work, at the same price[1][2].

Qwen3.8-Flash is the light multimodal model of the same generation, released in late August. As a Bilibili reviewer describes it, it is a mixture-of-experts model of 125B parameters with 6B active per call, and a native 256K context that extends to 1M[6].

Qwen3.8-Flash Qwen3.8-Max (0902)
List price per 1M tokens (input / output) ¥0.8 / ¥2.7 ¥12 / ¥36 ($2 / $6)
Context Up to 1M tokens 1M tokens
Input Text, images, video Text, images, video, audio
Model name on MoeRouter qwen3.8-flash qwen3.8-max-0902 (pinned to the 0902 build) and qwen3.8-max
Endpoints on MoeRouter Chat Completions, Responses, Messages Chat Completions, Responses, Messages, Gemini

Dollar prices are as MyDrivers and DataCamp report them[2][3]. Yuan prices and input types come from a user's post in the Alibaba Cloud developer community, which says its figures are taken from Alibaba Cloud's published rates[4].

One piece of context: Alibaba released the newer Qwen 4 in November 2026[9]. MoeRouter offers the Qwen3.8 line at present.

What they cost on MoeRouter#

per 1M tokens
ModelGroupInputOutputCache read
qwen3.8-flashdefaultCN¥0.32≈ US$0.048CN¥1.08≈ US$0.162CN¥0.064≈ US$0.0096
qwen3.8-max-0902defaultCN¥1.2≈ US$0.18CN¥3.6≈ US$0.54CN¥0.15≈ US$0.0225

Reference prices per 1M tokens in Chinese yuan (CNY), the currency accounts are charged in, from the model square (). Dynamically priced models settle at the route that served each request; the model square is authoritative. Amounts in USD are estimates, converted at 1 CNY = 0.15001 USD ().

Both are in the default group, which shows the reference price of the cheapest route at the moment; the usage log shows what each request was charged.

Benchmarks#

Qwen3.8-Max-0902#

Its headline is front-end coding. On CodeArena's front-end board (WebDev) the 0902 build gained 22 points to reach 1,691 and first place. On the same board's value chart its blended price is $5 per 1M tokens, against $20 and $12 for the models ranked second and third[1]. DataCamp gives the same board's other figures: Claude Opus 5 Max 1,687, Kimi K3 Max 1,674, the original Qwen3.8-Max 1,669[3].

Alibaba's own comparison table is broader and more sober[2][3]:

Benchmark 0902 build Original Qwen3.8-Max Claude Opus 5
TerminalBench 3.0 29.0 11.3 42.7
DeepSWE 1.1 69.3 56.6 73.6
QwenSWEBench V2 70.0 55.1 68.0
SWE-Marathon 44.8 39.1 50.0
JobBench 64.0 53.4 67.8
WorkArena (Elo) 1468 1348 1437

The 0902 build is far ahead of the original, yet it beats Claude Opus 5 on only 6 of the 14 benchmarks and trails on the other 8, TerminalBench 3.0 and DeepSWE 1.1 among them. Note that the comparison is with Opus 5, not the newer Opus 5.5. All of these figures are Alibaba's own; DataCamp's page has no independent results[3].

Qwen3.8-Flash#

On ReLE, NoneLinear's Chinese-language benchmark of about 15,000 questions, qwen3.8-flash scores 74.0% at about ¥10 per thousand calls; glm-5.3-flash, in the same class, scores 74.2% at ¥12.7. The two are level on overall ability in Chinese, with Qwen3.8-Flash slightly cheaper[5].

Hands-on#

AI产品狙击手 (Bilibili) ran his standing test suite on Qwen3.8-Flash in Claude Code with thinking at maximum. It passed all the reasoning items: constrained generation, the 24 game and a combination-lock puzzle. In the larger coding projects, a browser operating system, a space shooter and a Trello-style board, there were small bugs in layout, object orientation and button placement, and a 3D racing game came out the most finished. His summary: steady reasoning and creditable coding[6].

吉言2099 (Bilibili) tested Qwen3.8 Max three times in the ten days after it first appeared in late July, while it was still a preview. His record: mediocre on day one, somewhat stronger on day three, and keeping up with mainstream models by day ten. Alibaba had promised daily updates at the time[7]. A reminder to check the date of any early Qwen3.8-Max review: before and after the 0902 build are very different.

子昕 (Volcengine developer community) ran the same three real scenarios on several Chinese models. In the one with a real-time ticketing system and a waiting list, Qwen 3.8 Max got through the main flow of locking a seat, paying, issuing, refunding and the waiting list, where GLM-5.3-Flash stopped at a payment error[8].

Which to pick#

Good for

  • Flash: high-concurrency, cost-sensitive agents, reading images with text, code completion, prototypes
  • Max-0902: front ends and end-to-end app generation, complex multi-step tasks, long video and complex charts
  • Work that needs strong Chinese at a low price

Less suited to

  • Complex coding in a terminal: on Alibaba's own figures it is still well behind Claude on TerminalBench 3.0
  • Latency-sensitive use of Max: users report that Qwen overthinks, and a lower reasoning effort is worth trying
  • Wanting the newest generation: Qwen 4 is out

The post in the Alibaba Cloud developer community advises Max when results matter most and the budget allows, and Flash for multimodal work on a budget, at about a fifteenth of Max's price[4].

In the same price class, compare GLM-5.3-Flash and DeepSeek V4.1 Flash; for stronger coding, see Kimi K3.1 Pro or Claude Opus 5.5.

How to call them#

With the OpenAI SDK:

import os
from openai import OpenAI

client = OpenAI(
    base_url="https://moerouter.com/v1",
    api_key=os.environ["MOEROUTER_API_KEY"],
)
reply = client.chat.completions.create(
    model="qwen3.8-flash",
    messages=[{"role": "user", "content": "Rebuild the layout of this design in HTML and CSS"}],
)
print(reply.choices[0].message.content)

Both models answer on the Messages and Responses endpoints too, so they can be set as the model in Claude Code and Codex. Use qwen3.8-max-0902 when a project needs a fixed build and qwen3.8-max otherwise. The docs are in Chinese for now.

Frequently asked questions#

What is the difference between Qwen3.8-Flash and Qwen3.8-Max?#

Max is the flagship, with 2.4 trillion parameters, for complex multi-step tasks and end-to-end app generation. Flash is the light multimodal model at about a fifteenth of Max's list price, for high-concurrency and cost-sensitive work.

What is Qwen3.8-Max-0902?#

The upgrade of Qwen3.8-Max released on September 2, 2026: the same base model with post-training for coding and professional office work, at the same list price. It is first on Code Arena's front-end coding board with 1,691 points.

How much does the Qwen3.8 API cost?#

Qwen3.8-Max lists at $2 input and $6 output per 1M tokens, about ¥12 and ¥36. Qwen3.8-Flash lists at ¥0.8 input and ¥2.7 output. MoeRouter's reference prices are in the table above.

Is Qwen3.8-Max better than Claude?#

On the front-end coding board the 0902 build is ahead of Claude Opus 5 Max, 1,691 to 1,687. In Alibaba's own table of 14 benchmarks it trails Claude Opus 5 on 8, including terminal coding and DeepSWE, and the comparison is not with the newer Opus 5.5.

Qwen3.8-Flash or GLM-5.3-Flash?#

They are level on the Chinese benchmark, 74.0% to 74.2%. GLM-5.3-Flash has the lower reference price on MoeRouter. Both are cheap, so running your own tasks on each is the direct way to decide.

Is Qwen3.8 still worth using?#

Alibaba has released Qwen 4, but Qwen3.8 remains cheap and good enough: as of November 18, 2026, qwen3.8-flash was 9th by usage on MoeRouter over the previous 7 days. If you use it and are happy with the results, there is no rush to change.

Sources

The test results and figures quoted in this post come from the public sources below and belong to their authors; each quotation is marked where it appears.

  1. 阿里千问 Qwen3.8-Max-0902 版本模型发布,拿下前端编程总榜冠军 (in Chinese) IT Home (IT之家) · published 2026-09-02
  2. 千问Qwen 3.8 Max-0902版发布:1个月时间内代码能力暴涨 超越Opus5 (in Chinese) MyDrivers, republished by Sina Tech · published 2026-09-02
  3. Qwen3.8-Max: Features, Benchmarks, and Pricing DataCamp · published 2026-09-02
  4. 阿里云千问qwen3.8-max、qwen3.7-max、qwen3.8-flash、qwen3.7-flash热门大模型对比与选择参考 (in Chinese) Alibaba Cloud developer community (a user's post) · published 2026-09-04
  5. 【非线智能测评】智谱 GLM-5.3-Flash 实测:Coding与多模态大幅跃升,重塑轻量模型性价比标杆 (in Chinese) NoneLinear (非线智能) · published 2026-09-03
  6. Qwen 3.8 Flash测评报告!究竟有多能打? (video, in Chinese) Bilibili · AI产品狙击手 · published 2026-08-27
  7. Qwen3.8 Max发布第10天实测,和首发比可以说天壤之别! (video, in Chinese) Bilibili · 吉言2099 · published 2026-07-29
  8. 牛来原来是智谱,GLM-5.3-Flash 实测有惊喜也有硬伤 (in Chinese) Volcengine developer community · 子昕
  9. Qwen 4 正式发布!开源模型正面对决 Claude Fable 5.2,谁才是真正的最强 AI?实测效果来了!| 零度解说 (video, in Chinese) Bilibili · 零度解说 · published 2026-11-13

Disclosure: MoeRouter sells API access to the models discussed here. The test results in this post come from the third-party sources cited; prices and endpoint details come from MoeRouter's own data.