Qwen3.8 Max

Overview

Qwen3.8 Max is Alibaba's new flagship model, officially released on August 3, 2026. It has 2.4 trillion total parameters in a sparse MoE architecture (~95B active per request), with a hybrid attention mechanism (standard Multi-Head Attention + linear attention) that balances precision and efficiency across ultra-long contexts.

Compared to Qwen3.7 Max, the 3.8 Max upgrade spans three axes: native multimodal (image and video input join text), a larger output window (128K tokens, up from 65K), and long-horizon autonomous agent capability — in Alibaba's demonstration, the model ran continuously for 16 days, made 265 Git commits, and autonomously shipped a complete agent framework (Hermes Agent). Model weights are planned for open-source release around mid-August 2026.

Key capabilities

Qwen3.8 Max supports three cache modes (implicit / explicit / session). Implicit cache reads are priced at 12.5% of the standard input rate; explicit reads are lower (~8.5%); explicit cache creation is 125% of the input rate. Current pricing is available in the model catalog.

Benchmarks

Qwen3.8 Max evaluates across a broad flagship axis: coding and autonomous agent delivery, instruction following, research reproduction, and desktop operation — while introducing multimodal understanding to the Max tier for the first time. It ranks #2 on VisionArena and leads Fable 5 on IFBench and PaperBench.

Coding and agents

Coding has always been the Qwen Max series' strength. Qwen3.8 Max ranks #4 globally on CodeArena (Elo-based competitive coding leaderboard). The long-horizon agent evidence is even more striking:

• 16 days of continuous coding: 265 Git commits, autonomously shipping the complete Hermes Agent framework
• Competition performance: defeated 458 human teams in 24 hours on the Tianchi WWW2025 challenge
• Research reproduction: 125 hours of autonomous work, improving AIME24 scores by 2.7 points

Reasoning and instruction following

On hard reasoning, Qwen3.8 Max achieves 92.6 on GPQA Diamond, within 1 point of Opus 5 on the Arena blind leaderboard. More critically, it scores 82.8 on IFBench (instruction following), far ahead of Fable 5's 63.5 — meaning significantly higher instruction comprehension and execution accuracy in complex multi-step scenarios.

Multimodal and desktop operation

Qwen3.8 Max is the first Max-tier model with native visual understanding. It ranks #2 globally on VisionArena and scores 86.1 on OSWorld (desktop operation agent). Visual understanding spans the full planning-execution-verification loop, supporting screenshot OCR, PDF schematics, video temporal frames, and mixed-media parsing — none of which the text-only Qwen3.7 Max could do.

Research reproduction

PaperBench (research paper reproduction) scores 93.0, exceeding Fable 5's 88.8. The model can read papers, understand methodology, write experiment code, and iteratively tune parameters for end-to-end research assistance.

When to use it

• Long-horizon autonomous coding: end-to-end development of large projects, multi-day continuous operation with autonomous debugging and iterative delivery.
• Multimodal workflows: mixed text-image documents, video content understanding, screenshot OCR + code generation pipelines.
• Professional domain tasks: legal contract review, financial research reports, chip design optimization, and other deep-reasoning vertical scenarios.
• Agent orchestration: function calling + built-in tools + 262K thinking budget, suited for multi-step, long-context agent loops.

CrossModel exposes Qwen3.8 Max through an OpenAI-compatible API. Current pricing is available in the model catalog.