On July 19, 2026, Alibaba Cloud's Qwen team released Qwen3.8-Max Preview — with 2.4 trillion (2.4T) parameters, the largest flagship model in the Qwen series' history. The official line: its overall capability is second only to Anthropic's Fable 5, making it one of the "strongest models apart from Fable 5."

This is the first time a Chinese frontier model has moved its benchmark from GPT to the Fable 5 line. This article covers where it's strong, how to try it early, and when the API will be available.

What Qwen3.8 Is: The One-Sentence Positioning

Qwen3.8-Max is Qwen's flagship answer to Fable 5, focused on code engineering and professional office scenarios. The version released so far is a Preview; the official team says it "evolves on a day-by-day basis," with the stable release coming soon and open weights to follow.

Core Capabilities: Three Key Labels

1. 2.4T Parameters, the Largest in the Qwen Series

Parameter count reaches 2.4 trillion, the largest model in Qwen's history. In horizontal comparison, this is the biggest parameter scale among publicly available Chinese models.

2. Standout Code Engineering and Professional Office Performance

Official materials and the first wave of reviews point to the same conclusion: Qwen3.8-Max performs especially well in code engineering, long-horizon agent tasks, and professional office scenarios. With coding benchmarked against Fable 5, it suits:

• Large multi-file engineering refactors
• Long-horizon code agent tasks
• Complex data analysis and office automation
• Professional document drafting and rewriting

3. Improving "Day by Day"

The preview uses a continuous training strategy — the model weights aren't frozen — so every so often there's a perceptible capability jump. Even if your first impression is underwhelming, trying again a couple of days later may be a different story.

How to Try It Early: Three Entry Points

⚠️ Note: The preview's API is not yet fully open; for now, access is mainly through the products themselves. Don't go down the "apply for an API key and call the endpoint" path — the three entry points below are all direct-to-use.

Entry 1: Qwen's Official Website (Easiest, Free)

• URL: https://chat.qianwen.ai
• Select Qwen3.8-Max Preview in the model switcher menu
• Available directly on PC, free of charge

Best for: quickly testing quality and comparing everyday conversation

Entry 2: Token Plan Personal (Heavy Use)

• URL: https://platform.qianwenai.com
• Alibaba's newly launched Token Plan Personal subscription, with a limited-time offer starting at about 35–39 RMB/month
• Includes higher usage quotas for Qwen3.8-Max-Preview

Best for: heavy users who want it stable enough for daily work

Entry 3: Qoder / QoderWork (Developer Scenarios)

• Qoder (Alibaba's AI IDE): https://qoder.com
• QoderWork (a development platform for teams)
• Both products launched with day-one support for Qwen3.8-Max-Preview

Best for: writing code, engineering refactors, running long-horizon agent tasks

💡 Suggested path: try everyday conversation free on Qwen's website first, then decide whether to subscribe to Token Plan or test the coding ability inside Qoder. Don't pay upfront.

When Will the API Fully Open

The official roadmap:

• Now: Preview version, product-side experience only
• Soon: Stable release with open-sourced weights
• After the stable release: Full API access (DashScope on Model Studio)

So if you're a developer planning product integration, just wait for the stable release — even if preview API access exists, it's unstable and the weights will change; don't wire it into production.

How It Stacks Up Against Recent Frontier Models

Key differentiator: Qwen3.8-Max is the largest model in this wave and the only Chinese entrant explicitly benchmarked against Fable 5, with an official open-source commitment. If you're already in the Qwen ecosystem (Lingma, Qoder), it jumps to the top of your priority list.

Testing Suggestions: Try It With These Tasks

The preview doesn't come with published benchmarks, so testing it yourself is the most accurate approach. Three recommended test tasks:

Task 1: Multi-File Engineering Refactor (Tests Coding)

This is a mid-sized FastAPI project (~15 files).
Migrate all synchronous database access to async/await
while keeping every existing test passing. Output the refactor step by step.

Task 2: Long-Horizon Data Analysis (Tests Office Skills)

Based on this 200MB sales data CSV:
1. Find the SKUs whose repeat purchase rate has kept declining over the past 12 months
2. Identify the root cause of the decline (pricing / inventory / competitors)
3. Produce a 5-page analysis report for the CEO (Markdown)

Task 3: Complex Document Writing (Tests Reasoning and Expression)

Based on this 80-page industry research report,
write a 3,000-word policy impact analysis.
Every claim must cite a page number from the source; fabricating data is not allowed.

Compare the outputs of all three tasks against Fable 5, and you'll get a direct feel for what tier "second only to Fable 5" actually means.

FAQ

• Is the preview paid? The Qwen website and the basic Qoder experience are free; Token Plan is subscription-based (about 35–39 RMB/month).
• Can I deploy it locally? Not the preview. The official team says the stable release will be open-sourced, at which point local deployment becomes possible.
• How is its Chinese? The Qwen series has always been strong in native Chinese optimization, and Qwen3.8-Max sits in the top tier for long-form Chinese comprehension and handling of specialized terminology.
• Is the gap to Fable 5 large? The official self-assessment says "second only"; early hands-on testing generally agrees that coding is close, long-horizon agents slightly weaker, and general reasoning still trails — but the model is iterating day by day.

Related Links

• Qwen official website (free trial): https://chat.qianwen.ai
• Token Plan Personal: https://platform.qianwenai.com
• Qoder (developer IDE): https://qoder.com
• Coverage of Alibaba's Token Plan Personal launch: https://wallstreetcn.com/articles/3777367