Qwen3.8-Max

Quick Answer

Qwen3.8-Max is Alibaba’s current flagship model. It was previewed on July 19, 2026 at the World AI Conference in Shanghai and reached general availability on August 3, 2026. Alibaba states it has 2.4 trillion total parameters, uses a sparse Mixture-of-Experts architecture, is multimodal (text, image, video, documents), and supports a 1M-token context window. Standard API pricing is $2.00 per million input tokens and $6.00 per million output tokens.

What Changed Since Preview

The preview version launched with no model card, no active-parameter count, and no benchmark scores, an unusual gap for a flagship announcement. The GA release closed part of that gap: Alibaba published a benchmark table and standard pay-as-you-go pricing alongside the launch.

The open-weight release on August 13, 2026 closed most of the rest: it disclosed about 95 billion active parameters per token (out of 2.4 trillion total), which makes cost comparisons with other Mixture-of-Experts models like DeepSeek or GLM 5.2 possible. Alibaba’s benchmark comparisons with specific rival models are still its own reported numbers, so check independent evaluations before relying on them.

What Qwen3.8-Max Is Best For

• Multimodal tasks: text, image, video, and document input
• Long-context work: a 1M-token context window for full-repository or long-document tasks
• Coding and agent workflows: Alibaba reports strong results on agent-oriented benchmarks like Terminal-Bench, alongside weaker results on others like SWE-bench Pro, so performance varies by task type
• Watching the open-model landscape: a signal of where Alibaba’s Qwen line is heading, especially with open weights arriving soon

Key Facts (What’s Confirmed vs. Not)

Confirmed by Alibaba:

• Developer: Alibaba Cloud
• Previewed: July 19, 2026, at WAIC 2026 in Shanghai
• General availability: August 3, 2026
• Type: Multimodal, sparse Mixture-of-Experts model
• Total parameters: 2.4 trillion (stated by Alibaba)
• Context window: 1M tokens
• API pricing: $2.00 per million input tokens, $6.00 per million output tokens, via Alibaba Cloud Model Studio
• Active parameters: about 95 billion per token (disclosed with the open weights)
• Open weights: published August 13, 2026 as Qwen3.8-2.4T-A95B; text-only (the hosted model also accepts images and video)
• License: custom, not Apache 2.0. Model-as-a-service or AI-assistant businesses above $50 million in annual revenue need a separate commercial license, and very large products must display the model name

Check the model card for the full license before deploying commercially.

Why Parameter Count Alone Doesn’t Determine Quality

A 2.4-trillion-parameter total count sounds impressive, but for a Mixture-of-Experts model, what matters most for real-world cost and speed is the active-parameter count, how much of the model actually runs per token, not the total size. For Qwen3.8-Max it is about 95 billion, which is large for an active count: expect inference to cost more per token than MoE models with smaller active footprints, such as DeepSeek’s V4.1 Flash. Total parameters affect what the model can potentially know and do; active parameters affect what you actually pay and how fast you get a response.

Why Open Weights Don’t Mean Cheap Deployment

The weights are now public, but self-hosting is not free in any practical sense. A model with 2.4 trillion total parameters, even with a smaller active-parameter footprint per token, requires serious multi-GPU infrastructure to run at usable speed. Most individuals and small teams will still access it through Alibaba’s API rather than hosting it themselves, the same pattern seen with other large open-weight releases like GLM 5.2.

How to Access It

API access (available now): through Alibaba Cloud’s Model Studio, at $2.00 / $6.00 per million input/output tokens.

Self-hosted: the text-only weights are on Hugging Face as Qwen/Qwen3.8-2.4T-A95B, under a custom license. For most teams, the smaller sibling Qwen3.8-27B, already shipped as open weights on August 14, 2026, and is far more practical to self-host than the full 2.4T flagship, a 27-billion-parameter dense model built to run on a single consumer GPU.

Honest Limitations

• Large active footprint: about 95B active parameters per token makes it expensive to serve compared with leaner MoE models
• Benchmark scores are vendor-reported: check independent evaluations before relying on Alibaba’s comparisons
• Mixed results across benchmarks: strong on some (Terminal-Bench, PaperBench), weaker on others (SWE-bench Pro), so real-world fit depends on your specific task
• Open checkpoint is text-only, with a custom license: multimodal input needs the API, and larger commercial users need a separate agreement

Alternatives Worth Knowing

• Qwen3.8-27B, the smaller, open-weight, single-GPU sibling in the same model family
• GLM 5.3, an already fully documented open-weight model from Z.ai with published active-parameter counts and benchmarks
• Kimi K3, Moonshot AI’s open-weight frontier model
• DeepSeek, a widely used open-leaning model family with a track record of published specs, spanning cost-efficient V4.1 Flash and the larger V4-Pro
• Claude and ChatGPT, closed frontier alternatives with full documentation

For how to weigh a model like this against fully documented options, see Open Models vs Closed Models and Frontier Open Models Explained.