Qwen3.8-Max Explained: Specs, Pricing, and How It Compares (2026)

Estimated reading time: 10 minutes. Last verified: August 3, 2026.

Key takeaways

• Qwen3.8-Max is Alibaba’s new flagship in the Qwen family: a 2.4-trillion-parameter sparse MoE model that activates about 95 billion parameters per forward pass, with a 1-million-token context window.
• It shipped on August 3, 2026 via Alibaba Cloud Model Studio APIs and QwenWork (Alibaba’s workplace agent platform, which entered public beta the same day). Model weights are scheduled for open release the following week.
• Official Qwen Cloud list price is $2 per 1M input tokens and $6 per 1M output tokens, with aggressive context-cache rates ($0.25 / $0.17 per 1M on implicit / explicit cache read). That undercuts the previous Qwen3.7-Max list price ($2.50 / $7.50).
• Alibaba cites 5th on Text Arena, 2nd on Vision Arena, and 4th on Frontend Code Arena. Treat those as vendor-reported leaderboard positions, not third-party audited scores.
• For builders: this is a long-horizon coding and multimodal-agent model first, a cheaper chat bot second. If raw API cost is the only constraint, DeepSeek V4 still undercuts it by a wide margin.

Table of contents

1. What Qwen3.8-Max is
2. Specs that matter
3. What it is built to do
4. Pricing and access
5. How it compares
6. Who should care
7. FAQ

---

What Qwen3.8-Max is

Model release names get noisy fast. "Qwen3.8-Max" is not a new company, and it is not a small research checkpoint. It is Alibaba’s top hosted model in the Qwen line as of August 3, 2026.

For practitioners, the useful definition is simple:

Qwen3.8-Max is a multimodal Mixture-of-Experts language model (2.4T total parameters, ~95B active) with a 1M-token context window, sold through Alibaba Cloud’s Model Studio API and used inside QwenWork, with open weights promised shortly after launch.

That definition separates three things people often mix up:

• Qwen (the product family): consumer Studio apps plus the Cloud API catalog. See the Qwen AI tool page.
• Qwen3.8-Max (this model): the current Max-tier flagship in that catalog.
• Open weights (coming next): a downloadable checkpoint Alibaba says will ship the week after the API launch. Until that lands with a license and Hugging Face (or equivalent) repo, "open-source Max" is a roadmap claim, not a downloadable artifact.

A simple way to think about it: Max is the paid frontier tier. Flash and Plus stay the cost/latency tiers underneath it.

---

Specs that matter

Verified against Alibaba’s August 3, 2026 press release and the live Qwen Cloud model page.

Spec	Qwen3.8-Max
Total parameters	2.4 trillion (sparse MoE)
Active parameters	~95 billion
Architecture	Sparse MoE + hybrid attention (built on the Qwen 3.5 foundation)
Context window	1M tokens
Max input / output	991K / 131K tokens (983K / 131K in thinking mode)
Modalities (input)	Text, image, video
Modalities (output)	Text
Leaderboard claims (Alibaba)	Text Arena #5, Vision Arena #2, Frontend Code Arena #4
Hosted access	Alibaba Cloud Model Studio API, QwenWork
Open weights	Scheduled for the week after Aug 3, 2026
API model id	qwen3.8-max

The MoE detail is the practical one. A 2.4T dense model would be brutal to serve. Activating ~95B parameters is how Alibaba can claim frontier-scale capacity without dense-model inference cost. Cache pricing on the Cloud page is another signal that long-context and agent loops are the intended workload, not one-shot chat.

---

What it is built to do

Alibaba’s pitch is not "slightly better chat." It is long-horizon agent work with vision in the loop.

Autonomous coding and long projects

Press materials describe multi-day autonomous coding: generating, testing, previewing, reading logs, and iterating without constant human prompting. Alibaba’s showcase is oh-my-cli, a self-evolving agent framework the model supposedly built over a 16-day internal run and later open-sourced on GitHub. Marketing language aside, the product intent is clear: Max is positioned for agent frameworks and multi-day engineering loops, not just autocomplete.

Multimodal recreation and visual agents

The model accepts images and video, not only text. Documented demo classes include:

• rebuilding frontend apps from a UI screenshot
• turning 2D floor plans into 3D interior views
• editing raw footage into polished vlogs
• generating educational animations and interactive games from natural language
• ingesting long documents, TV series, or long livestreams into searchable knowledge bases

Alibaba also introduced RecreationBench, a black-box app-recreation benchmark where the model rebuilds applications from interaction and visual feedback only (no source code or internet). Treat that as a vendor benchmark until independent reproductions show up.

Professional and office workloads

Coverage spans legal review, financial research, design, sports analytics, and similar multi-constraint office tasks. Same day as the model launch, QwenWork entered public beta as Alibaba’s workplace agent product, competing with tools in the same lane as Claude Cowork-style products and ChatGPT work surfaces, plus Chinese rivals like Kimi Work.

What that means in practice:

• What it is: a multimodal Max-tier model aimed at coding agents, long context, and office automation.
• What limits it: hosted access still sits inside Alibaba Cloud’s ecosystem; compliance, data residency, and region choice matter. Open weights are promised, not yet shipped as of verification day.
• What it is not: a drop-in replacement for every cheaper open model. If you only need solid coding chat at the lowest $/token, look at DeepSeek first.

---

Pricing and access

Hosted API (verified August 3, 2026)

From qwencloud.com/models/qwen3.8-max:

Meter	Price
Input	$2.00 per 1M tokens
Output	$6.00 per 1M tokens
Input (implicit cache)	$0.25 per 1M tokens
Explicit cache creation	$2.50 per 1M tokens
Explicit cache read	$0.17 per 1M tokens
Rate limits (listed)	2M TPM, 15K RPM

Compared with the previous Max tier in our directory notes (Qwen3.7-Max at $2.50 / $7.50), list prices moved down, not up. Cache discounts matter a lot for agent loops that re-send the same system prompt and repo context every turn.

API surface supports DashScope and OpenAI-compatible calling patterns. Features called out on the model page: function calling, structured outputs, web search, prefix completion, batches, and context cache. Fine-tuning is listed as a platform capability; confirm availability for this specific model id before planning a fine-tune budget.

Consumer / workplace access

• Qwen Studio / qwen.ai: consumer assistant surface for the broader Qwen family.
• QwenWork: workplace AI agent platform (public beta as of Aug 3, 2026), web and desktop.

Token Plan subscription packaging still exists for heavier agent/coding products in the Qwen Cloud lineup. If you are budgeting from a credit plan rather than raw pay-as-you-go, re-check the live Token Plan page; promotional credit burn rates change often.

Open weights

Alibaba says weights ship the week after the API launch. SCMP frames this as a return to open-sourcing top-tier Qwen models after keeping several earlier 2026 flagships proprietary. Some Chinese coverage also mentions a companion Qwen3.8-27B open release. Until files, license, and hardware requirements are public, plan for hosted API use and treat self-hosting as pending.

---

How it compares

	Qwen3.8-Max	Qwen3.7-Max (prior Max)	DeepSeek-V4-Pro	Claude / ChatGPT flagships
Role	Current Qwen Max flagship	Previous Max	Cost/performance open-weight + API rival	US lab closed flagships
Parameters	2.4T MoE (~95B active)	Not disclosed the same way	Lab-disclosed V4 family (not 2.4T MoE)	Closed
Context	1M	Up to 1M (family)	1M on V4 tiers	Varies by model/product
API list price (approx.)	$2 / $6 per 1M in/out	$2.50 / $7.50	$0.435 / $0.87	Higher on frontier tiers; product seats separate
Open weights	Promised next week	Prior Max tiers often closed	Yes (MIT-style open weights on current V4 line)	No
Multimodal	Text + image + video in	Strong multimodal family	Strong text/coding focus; check current vision SKUs	Strong, productized
Best fit	Long-horizon agents, visual coding, Alibaba Cloud stacks	Teams already on 3.7 Max	Lowest $/token frontier coding	Enterprise workflow, US compliance defaults

Sources for competitor prices: DeepSeek pricing notes verified August 2026; prior Qwen Max from our July 2026 Qwen AI research; Claude/ChatGPT from their consumer/API pricing pages (seat plans and token APIs are different products).

Rough decision guide:

• Pick Qwen3.8-Max if you want Max-tier multimodal agents, 1M context, and either Alibaba Cloud hosting or (soon) open weights at this scale.
• Pick DeepSeek if price-per-token dominates and you are fine with a different product surface.
• Pick Claude or ChatGPT if your constraint is US-vendor procurement, existing workspace tooling, or compliance defaults, not raw model scale.
• Use Together AI (together-ai) or similar hosts when you want open Qwen-family checkpoints on third-party inference after weights land, without running your own cluster.

---

Who should care

Developers shipping coding agents. The long-horizon coding story and Frontend Code Arena claim are the core reason to trial qwen3.8-max now, especially if you already speak OpenAI-compatible APIs.

Teams drowning in long multimodal context. Hundred-page docs, long video, and screenshot-to-app loops are the multimodal pitch. Cache pricing makes repeated long prompts less painful than they look at sticker rates.

Builders waiting on open Max weights. If your plan is self-host or fine-tune, watch the open-weights drop next week and verify license + VRAM needs before rewriting infra. Do not block a product launch on a promised checkpoint that is not public yet.

Who can wait. If you only need cheap general chat or light coding assist, Flash/Plus tiers (or DeepSeek V4-Flash) are still the rational default. Max is for hard, multi-step work.

Honest tradeoff: Qwen3.8-Max is strong on scale, multimodal agents, and open-weight ambition. It is weaker as a "set and forget" choice for teams that need non-Alibaba data residency without careful region setup, or for buyers who refuse any model whose top Arena numbers are still mostly vendor-announced.

---

FAQ

What is Qwen3.8-Max?

Qwen3.8-Max is Alibaba’s August 2026 flagship model in the Qwen family: a 2.4-trillion-parameter sparse MoE multimodal model with about 95 billion active parameters and a 1-million-token context window. It is available through Alibaba Cloud Model Studio APIs and QwenWork, with open weights scheduled shortly after launch.

How much does Qwen3.8-Max cost?

On Qwen Cloud’s model page (verified August 3, 2026), pay-as-you-go pricing is $2 per 1M input tokens and $6 per 1M output tokens. Implicit cache input is $0.25 per 1M tokens; explicit cache read is $0.17 per 1M tokens. Subscription Token Plans may bill in credits instead of raw tokens depending on product.

Is Qwen3.8-Max open source?

Not on launch day. Alibaba announced that model weights will be released the week after the August 3, 2026 API launch. Until the files and license are public, treat it as a hosted model with an open-weights commitment.

How does Qwen3.8-Max compare to DeepSeek?

Qwen3.8-Max is larger on paper (2.4T MoE) and pushes multimodal / long-horizon agent demos harder. DeepSeek’s current V4 API prices are much lower (V4-Pro around $0.435 / $0.87 per 1M tokens as of August 2026). Choose Qwen for multimodal Max-tier agents and impending open Max weights; choose DeepSeek when cost-per-token and already-open weights matter more. See DeepSeek and Qwen AI.

How do I access Qwen3.8-Max?

Use the qwen3.8-max model id on Alibaba Cloud Model Studio / Qwen Cloud (DashScope or OpenAI-compatible clients), or try it inside QwenWork’s public beta. Consumer chat for the broader family still lives on qwen.ai.

Does Qwen3.8-Max replace Qwen3.7-Max?

For new Max-tier work, yes: 3.8-Max is the current flagship with lower list prices than 3.7-Max in our last verified notes. Existing integrations on qwen3.7-max should plan a model-id migration and re-run evals; do not assume identical behavior on coding or vision tasks.

---

Sources

• Alibaba Cloud press release, Aug 3, 2026
• Qwen Cloud model page: qwen3.8-max
• SCMP coverage, Aug 3, 2026
• Internal tool pages: Qwen AI, DeepSeek, Claude, ChatGPT, Kimi AI, Together AI