Qwen3.8-Max-Preview

Overview

Alibaba previewed Qwen3.8-Max-Preview on July 19, 2026, at the World Artificial Intelligence Conference (WAIC) in Shanghai, positioning it as the successor to Qwen3.7-Max. At 2.4 trillion total parameters built on a sparse Mixture-of-Experts architecture, it's the first Qwen model to cross the 1-trillion-parameter mark, and the first to do it as a multimodal system that handles text, images, video, and documents in a single model rather than a separate vision variant.

The timing is the real story. Qwen3.8-Max-Preview arrived three days after Moonshot AI shipped Kimi K3, a 2.8 trillion-parameter model that launched with full open weights and a published evaluation table. Alibaba's response leads with a bigger competitive claim, second only to Anthropic's Claude Fable 5, but backs it with nothing that can be independently checked. No benchmark names, no scores, no methodology, no third-party listing on Artificial Analysis or LMArena. For a company that has historically shipped detailed benchmark tables alongside its Max-series launches, that's a conspicuous omission, and one worth flagging plainly rather than repeating the claim at face value.

This is also, by name, a preview. Alibaba has said a fully open-weight release of the flagship Qwen3.8 is coming, following the pattern set by Qwen3.6-Max-Preview before it converted to a closed-weight production release. No date, license, or Hugging Face repository has been confirmed. Until that happens, Qwen3.8-Max-Preview exists only as a closed, API-gated build accessible through a handful of Alibaba's own platforms.

Coverage of the July 19 preview announcement, framed explicitly against Kimi K3's open-weight launch three days earlier. Source: marktechpost.com

Key Specifications

Benchmark Performance

There is no benchmark table to build here, and pretending otherwise would violate the basic premise of this page. As of this writing, Alibaba hasn't published a technical report, a model card, a single named benchmark score, or a methodology note for Qwen3.8-Max-Preview. No third-party tracker, not Artificial Analysis, not LMArena, not OpenRouter, has an entry for it yet.

The predecessor and competitor columns above exist only to show what a real comparison would need, not to imply anything about Qwen3.8-Max-Preview's actual standing. Alibaba's own framing is that the new model "should beat Qwen3.7-Max on coding, full-stack development, data analysis, and office workflows," which is a directional claim, not a measured one. On the coding benchmarks leaderboard, Qwen3.7-Max's verified numbers are what's currently tracked for the Qwen line; Qwen3.8-Max-Preview isn't on it because there's nothing to enter.

The "second only to Fable 5" line is the part to examine hardest. Alibaba hasn't disclosed which tasks were compared, which Fable 5 configuration was used, whether thinking mode was enabled on either model, or what the actual score gap was. Compare that to Kimi K3's launch three days earlier, which shipped with a published LMArena result (#1 on the Frontend Code Arena) and an Artificial Analysis Coding Index entry that anyone can check. Qwen3.8-Max-Preview offered marketing language in place of a table. Total parameter count, 2.4 trillion, also isn't a meaningful standalone comparison point without the active-parameter figure that determines actual inference cost and, historically, correlates more closely with real-world capability than total weight count does.

None of this means the underlying model is weak. It means the claim is currently unverifiable, and unverifiable claims from a lab with a documented interest in the outcome belong in the "wait and see" category, not the "confirmed" one.

Key Capabilities

Native multimodality at scale. This is Qwen's first model above 1 trillion parameters to handle images, video, and documents natively rather than through a bolted-on vision encoder or a separate model variant. Prior Qwen Max releases, including Qwen3.7-Max, shipped text-only. If the multimodal claims hold up under independent testing, it closes a real gap against GPT-5.5 and Gemini 3.5 Pro, both of which have offered native multimodal input for several generations.

Thinking mode with tool orchestration. The model supports an explicit thinking mode with function calling and built-in tools, following the pattern Alibaba established with Qwen3-Max-Thinking: the model reasons about which tool to invoke and how to interpret results before responding, rather than calling tools reactively. Combined with native OpenAI and Anthropic Messages API compatibility (a carryover from Qwen3.7-Max), teams already running agent harnesses against either protocol can point them at Qwen3.8-Max-Preview with a base URL change.

Long-horizon office and coding workflows. Alibaba specifically calls out "full-stack development, data analysis, and Office workflows" as the areas where it expects Qwen3.8-Max-Preview to beat Qwen3.7-Max, targeting the kind of long-running, multi-step agentic tasks Qwen3.7-Max was already built around. The 1M-token context window carries over unchanged from the prior generation rather than expanding further.

Qoder's own event page frames the preview mainly around its temporary pricing discount rather than technical specifications. Source: qoder.com

Pricing and Availability

Qwen3.8-Max-Preview is live now, but only through three Alibaba-controlled surfaces: the Token Plan subscription service, Qoder, and QoderWork. There's no standalone per-token API price published anywhere. What exists instead is a preview discount: 10% of the standard rate across the board, which implies a standard rate that hasn't itself been disclosed.

Token Plan, Alibaba's subscription tier for individual and small-team access, runs $6/month for Lite, $20/month for Standard, and $70/month for Pro internationally (39/139/499 CNY domestically), credit-metered rather than priced per million tokens. Qoder's own developer terms list Qwen3.8-Max-Preview at a 0.5x credit multiplier against standard-rate models, dropping to 0.05x (90% off) during regular hours and 0.01x (98% off) during off-peak hours (22:00 to 08:00 SGT). None of this maps cleanly onto the dollar-per-million-token pricing used to compare Qwen3.7-Max ($2.50/$7.50) or Kimi K3 ($3.00/$15.00) against the rest of the field, which makes direct cost comparison impossible until Alibaba publishes a real rate card.

There's no self-hosting option today. The promised open-weight release of the full Qwen3.8 has no confirmed date, license, or Hugging Face listing. Until it ships, closed API access through the three listed platforms is the only way to use this model, and even that access is explicitly framed by Alibaba as a preview subject to change before a production release.

Strengths and Weaknesses

Strengths

• First Qwen Max-line model with native multimodal input (text, image, video, document) above 1 trillion parameters
• 1M-token context window carried over from Qwen3.7-Max, with thinking mode, function calling, and built-in tools
• Native OpenAI and Anthropic Messages API compatibility eases migration from existing agent harnesses
• 2.4 trillion total parameters signals continued aggressive scale investment from Alibaba just three days after Kimi K3's 2.8T launch

Weaknesses

• No model card, no benchmark table, no active-parameter-per-token disclosure as of this writing
• "Second only to Fable 5" is Alibaba's own unverified claim, with no named benchmarks, no methodology, and no third-party corroboration
• No confirmed API pricing; "10% of standard rate" is meaningless without a published standard rate
• Closed, preview-only access via Token Plan, Qoder, and QoderWork; no self-hosting path
• Open-weight release promised "soon" with no date, following Qwen3.6-Max-Preview's precedent of eventually shipping closed instead

Related Coverage

• Qwen3.7-Max model profile - the direct predecessor and current production Qwen flagship
• Qwen3.6-Max-Preview model profile - the last Qwen preview build, which later shipped closed-weight
• Kimi K3 model profile - Moonshot AI's 2.8T open-weight model that prompted this preview's timing
• Claude Fable 5 model profile - the model Alibaba claims to trail as "second only to"
• Qwen 3.6 Max review - hands-on testing of the prior Qwen preview generation
• Coding benchmarks leaderboard - where Qwen3.7-Max's verified scores currently sit
• Open source LLM leaderboard - pending Qwen3.8's promised open-weight release
• Alibaba shifts Qwen to closed weights with Qwen3.6-Max - the precedent for this preview eventually going closed

FAQ

Is Qwen3.8-Max-Preview open source?

No, not yet. It's currently closed-weight and API-only, accessible through Alibaba's Token Plan, Qoder, and QoderWork. Alibaba has promised open weights for the full Qwen3.8 "soon," with no confirmed date or license.

How many parameters does Qwen3.8-Max-Preview have active per token?

Alibaba hasn't disclosed this. The model has 2.4 trillion total parameters on a sparse Mixture-of-Experts architecture, but the active-parameter count that determines real inference cost and latency is unpublished.

Is Qwen3.8-Max-Preview really "second only to Fable 5"?

That's Alibaba's own unverified claim. No benchmark names, scores, or methodology have been published, and no independent evaluator has scored the model yet. Treat it as a marketing position until third-party data appears.

How much does Qwen3.8-Max-Preview cost to use?

There's no standalone per-token price. Access runs through Alibaba's Token Plan subscription ($6 to $70/month) or Qoder/QoderWork credit multipliers, both discounted to roughly 10% of an unpublished standard rate during the preview period.

How does Qwen3.8-Max-Preview compare to Kimi K3?

Directly, it's hard to say. Kimi K3 shipped with a published LMArena ranking and Artificial Analysis benchmark entries; Qwen3.8-Max-Preview has neither. Kimi K3 is also already available as open-weight-in-progress with a stated release date for full weights, while Qwen3.8-Max-Preview remains fully closed with no such date.

Does Qwen3.8-Max-Preview support images and video?

Yes. Alibaba describes it as processing text, images, video, and documents, making it the first Qwen Max-series model above 1 trillion parameters with native multimodal input. This hasn't been independently verified in detail yet.

Sources:

• Alibaba previews Qwen3.8, claiming its strength trails only Anthropic's Fable 5 - South China Morning Post
• Alibaba Previews Qwen3.8-Max, a 2.4 Trillion-Parameter Multimodal Model, Days After Moonshot's Kimi K3 Open-Weight Launch - MarkTechPost
• Alibaba's Qwen takes on Kimi K3 with open-weight Qwen 3.8, says model is "second only to Fable 5" - The Decoder
• Qwen3.8: 2.4T Parameters, Open Weights, No Benchmarks - Techsy
• Is Qwen3.8 Really "Second Only to Fable 5"? The Contents of the 2.4-Trillion-Parameter Model Remain Unseen - XenoSpectrum
• Qwen3.8-Max-Preview Launch - 90 Percent Off, Night-Time 98 Percent Off - Qoder Docs
• Qwen 3.8 Max Preview - Token Plan Guide - ExplainX