What is Qwen3.8-Max?
Qwen3.8-Max is the top model in Alibaba's Qwen3.8 family and the first Qwen Max model with open weights. It is a 2.4T-parameter mixture-of-experts model with 95B parameters active per token. The hosted version reads text, images and video across 1M tokens, thinks by default at three effort levels, and plugs into Claude Code, Codex, Qwen Code and OpenClaw.
The download, Qwen3.8-2.4T-A95B, is a narrower build: text only, always thinking, and about 4.9 TB at full precision. Alibaba recommends Qwen3.8-Max for its strongest reasoning and now lists Qwen3.7-Max as legacy.
What changed from Qwen3.7-Max?
Higher scores across Alibaba's table, image and video input, open weights and a lower price on Qwen Cloud.
- Coding
- Terminal Bench 2.1 rises from 74.5 to 86.6 and SWE-bench Pro from 60.6 to 67.7 in Alibaba's table.
- Agents
- Toolathlon Verified rises from 49.7 to 72.5, and CoWorkBench from 64.6 to 74.8.
- Price
- $2 / $6 per million tokens on Qwen Cloud, down from $2.50 / $7.50 for Qwen3.7-Max.
- Status
- Alibaba now lists Qwen3.7-Max as a legacy model and points to Qwen3.8-Max for the strongest reasoning.
Who is Qwen3.8-Max best for?
Qwen3.8-Max is best for coding agents in Claude Code or Codex, long unattended runs, document-heavy office work, and teams that want near-frontier results at a lower token price.
Coding agents in Claude Code or Codex
Qwen Cloud speaks the Anthropic and OpenAI Responses formats, and Alibaba's table puts it at 86.6 on Terminal Bench 2.1, ahead of Claude Opus 4.8's 84.6.
Long unattended runs
Alibaba shows it working alone for more than 10 days on a coding project and about 125 hours reproducing, then improving, a research paper.
Document-heavy office work
It scores 74.8 on Alibaba's CoWorkBench of legal, finance and medical tasks, and the hosted model reads 200-page PDFs and long videos.
Teams watching token prices
At $2 / $6 per million tokens, or $1.65 / $4.951 on Model Studio's Global scope, it costs far less per token than Claude Opus 5.5 at $4 / $20.
What's included with Qwen3.8-Max?
Three reasoning effort levels, thinking kept across turns, image and video input on the API, built-in web and code tools, cheap cache hits, and open weights.
- Three reasoning effort levels
- xhigh by default, plus medium and low; the hosted model can also answer without thinking.
- Preserved thinking
- Reasoning from earlier turns stays in context by default, which Qwen recommends for agent work.
- Image and video input
- The hosted model reads screenshots, charts, long PDFs and video; the open weights read text only.
- Built-in tools
- Web search, web extraction, code interpreter and image search on the Responses API, plus web fetch and search on the Anthropic-style API.
- Cheap cache hits and batch jobs
- Cached input costs $0.25 per million tokens, and batch calls on Qwen Cloud cost half the real-time price.
- Open weights
- Qwen3.8-2.4T-A95B comes in BF16 and FP8 for vLLM, SGLang and TokenSpeed.
Are Qwen3.8-Max's open weights the same as the API model?
No. The download, Qwen3.8-2.4T-A95B, is the base that the hosted Qwen3.8-Max builds on, with fewer features.
- Inputs
- The weights read text only; the API also takes images and video.
- Thinking
- The weights always think; the API can also answer without thinking.
- Context
- The weights run 262,144 tokens natively; the API serves 1M tokens by default, with built-in tools.
- Newer snapshot
- Qwen has not released weights for the qwen3.8-max-0902 update.
How much does Qwen3.8-Max cost?
Qwen3.8-Max is free in Qwen Chat, costs $2 per million input tokens and $6 per million output tokens on Qwen Cloud, and is in Qwen's Token Plan from $8 a month; the text-only weights download free.
- Qwen says Qwen Chat is free to use
- Offers Qwen3.8-Max with a 1M-token context
- Web, iOS, Android, macOS and Windows apps
- Qwen3.8-2.4T-A95B on Hugging Face, in BF16 or FP8
- Text only and always thinking; about 4.9 TB in BF16
- Extra licence terms only for very large businesses
- Cached input $0.25; batch calls half price
- 1M-token context, up to 131K tokens out
- 1M free tokens for new accounts, for about 90 days
- Lite: 11,500 credits a month
- Essential $16, Standard $25, Pro $80 a month
- Credits also cover Qwen3.7-Plus and other models
Qwen Cloud is Qwen's official API and lists prices in US dollars. Alibaba Cloud Model Studio charges $1.65 / $4.951 per million tokens on its Global scope in US (Virginia), Frankfurt, Hong Kong and Tokyo, and $2 / $6 in Singapore; it has no US-only scope for this model. Token Plan credits per request depend on the model, tokens, thinking and tools, a Credit Pack adds 20,000 credits for $15, and plans are not refundable. Team seats start at $30 per seat a month. Full pricing on qwencloud.com. We checked these prices on .
How do readers rate Qwen3.8-Max?
No one has rated Qwen3.8-Max yet. Be the first.
Sign in to rate. One rating per account.
Qwen3.8-Max pros and cons
It comes close to Claude Opus and Fable on coding for a much lower token price, but it burns many reasoning tokens, and its open weights are huge, text only and under a custom licence.
Pros
- Hands-on users report it finishing real coding work on par with Claude Opus and Fable models, in some cases fixing projects that Opus 4.8 had struggled with.
- At $2 / $6 per million tokens it costs less per token than Kimi K3 or Claude Opus 5.5, and it is free to try in Qwen Chat.
- Alibaba's own table puts it ahead of Claude Opus 4.8 on Terminal Bench 2.1 (86.6 against 84.6), PaperBench (93.0 against 80.3) and IFBench (82.8 against 62.2).
- It is the first Qwen Max model with open weights, and its licence allows commercial use with extra terms only for very large services.
Cons
- It spends many reasoning tokens: several Hacker News users note that its cost per task can match or beat pricier models despite the low token price, and one widely shared test drawing took 11 minutes.
- The open weights are about 4.9 TB in BF16, so self-hosting takes a data-centre server; users call it out of reach for local use.
- The open weights are not the API model: they read text only, cannot turn thinking off and run 262K tokens natively, which several users called a letdown.
- The licence is not MIT or Apache: model hosts and coding or office assistant businesses with more than $50M a year in revenue need a separate licence from Qwen.
What are the best Qwen3.8-Max alternatives?
Kimi K3 for open weights that read images, GLM-5.3 for cheaper tokens and security work, and Claude Opus 5.5 for a US provider and a Claude subscription.
Kimi K3FreeChoose it when you need open weights that read images: Moonshot's 2.8T model takes images as well as text in the weights you download, though its API costs more at $3 / $15 per million tokens.
GLM-5.3FreeChoose it when price per token and security work matter most: Z.ai charges $1.4 / $4.4 per million tokens and reports a leading CyberGym score, though it reads text only.
Claude Opus 5.5PaidChoose it when you need a US provider or want the model inside a Claude subscription: it is on Claude Pro at $20 a month, or $4 / $20 per million tokens on the API.
Questions about Qwen3.8-Max
What is qwen3.8-max-0902?
It is an upgraded snapshot of Qwen3.8-Max that Qwen says improves coding, multi-tool agent work and image understanding. It keeps the 1M-token context and the same per-token price on Alibaba's Model Studio, and Qwen has not released weights for it.
Who is Qwen3.8-Max for?
Qwen3.8-Max is best for coding agents in Claude Code or Codex, long unattended runs, document-heavy office work, and teams that want near-frontier results at a lower token price.
Is Qwen3.8-Max free?
Yes. Qwen3.8-Max has a free Qwen Chat plan: qwen says Qwen Chat is free to use.
How much does Qwen3.8-Max cost?
Qwen3.8-Max's paid plans are Qwen Cloud API at $2 / $6 per million input / output tokens and Token Plan at From $8 per month.
What are the best Qwen3.8-Max alternatives?
The closest alternatives on Recaply are Kimi K3, GLM-5.3 and Claude Opus 5.5.
Where can I try Qwen3.8-Max?
On its official website, qwen.ai.
Can I run Qwen3.8-Max on my own hardware?
Only on a large server. The open weights, Qwen3.8-2.4T-A95B, take about 4.9 TB in BF16 and 2.5 TB in FP8, and Qwen serves them with vLLM, SGLang or TokenSpeed. They read text only and always think, unlike the hosted Qwen3.8-Max.
Can I use Qwen3.8-Max commercially?
Yes. The Qwen3.8-Max License allows commercial use, changes and hosting. Products with more than 100 million monthly users or $20 million in monthly revenue must show the model name, and model hosts or coding and office assistant businesses above $50 million a year need a separate licence.
Can I use Qwen3.8-Max in Claude Code or Codex?
Yes. Qwen Cloud accepts the Anthropic API format for Claude Code and the OpenAI Responses format for Codex; set the model to qwen3.8-max and use your API key. Qwen Code, Qoder and OpenClaw work too, and the Token Plan covers these tools.
Our verdict
Qwen3.8-Max is the Qwen model to use through an API, and a real alternative to Claude Opus for coding agents.
Alibaba's tests put it level with Opus 4.8 on terminal coding at a fraction of the token price, and users report it holding up on real projects.
- Skip it for
- Anyone who plans to self-host it, since the weights are huge and text only, and teams that cannot send data to Alibaba's cloud.
- Check first
- Measure cost per task on your own work, because its long reasoning can cancel out the low token price.
Qwen3.5 Small
Qwen3.6-Max-Preview
Qwen3.7-Max
Qwen3.7-Plus
Qwen3.8-Flash
Qwen3.6-Plus