Alibaba Drops Qwen 3.6 Max Preview—Its Most Highly effective Mannequin But – Decrypt

Alibaba launched a preview of its subsequent flagship AI mannequin on Monday, pushing the Qwen sequence additional into the frontier race. Qwen 3.6-Max-Preview is probably the most highly effective mannequin the corporate has shipped to this point, topping six main coding benchmarks and posting significant positive factors in world information and instruction following over its predecessor, Qwen 3.6-Plus.

The mannequin is on the market now by Qwen Studio and the Alibaba Cloud Mannequin Studio API below the string qwen3.6-max-preview. It’s a proprietary, hosted mannequin with no open weights, and its API is appropriate with each OpenAI and Anthropic specs, which means builders can plug it into present pipelines with minimal adjustments.

That is additionally a shift in Alibaba’s enterprise mannequin, because it was recognized to offer highly effective open-source fashions by default. The decrease finish fashions are nonetheless open supply.

In line with Qwen’s official weblog, Qwen 3.6-Max-Preview ranked first throughout a number of main benchmarks that check coding and agent capabilities, together with SWE-bench Professional (real-world software program engineering duties), Terminal-Bench 2.0 (command-line execution), SkillsBench (normal problem-solving), QwenClawBench (software use), QwenWebBench (internet interplay), and SciCode (scientific programming).

By way of positive factors over Qwen 3.6-Plus, benchmarks for agentic expertise put it on prime of the household and different fashions like Calude 4.5 or GLM 5.1. The mannequin additionally improved in information benchmarks, with SuperGPQA (superior reasoning) rising by 2.3% and QwenChineseBench (Chinese language language efficiency) by 5.3%. Instruction-following capability, measured by ToolcallFormatIFBench, put it on prime of the rankings, beating Claude.

The discharge lands three days after Alibaba open-sourced Qwen 3.6-35B-A3B, a 35-billion-parameter mannequin that prompts solely 3 billion parameters per inference.

That strategy is designed to chop compute prices with out sacrificing output high quality. Mixed with Monday’s Max launch, the Qwen 3.6 lineup now spans Max-Preview on the prime of the household, the Qwen Plus variance for balanced workloads, Flash for speed-first duties, and 35B-A3B for these working domestically.

The Max-Preview additionally ships with preserve_thinking, a function that carries reasoning traces throughout multi-turn conversations. Alibaba particularly recommends it for agentic duties the place context continuity issues. For builders working autonomous brokers or long-running code era workflows, that could be a significant addition.

As Decrypt reported final week, Alibaba shut down the free tier of Qwen Code simply days after fellow Chinese language lab MiniMax rewrote its open-source license to dam industrial use with out written authorization. Each strikes sign a broader shift: Chinese language AI labs that constructed huge adoption on free, open companies at the moment are pivoting towards monetized, proprietary choices. Qwen overtook Meta’s Llama as probably the most deployed self-hosted mannequin on the planet—and that momentum was constructed nearly fully on free entry.

That free-to-paid transition runs parallel to a different pattern. Chinese language open fashions went from 1.2% of world open-model utilization in late 2024 to roughly 30% by finish of 2025, with Qwen main the cost. The Qwen 3.6-Max-Preview is the proprietary tip of that spear because it’s the mannequin Alibaba is betting will compete immediately with OpenAI’s GPT and Anthropic’s Claude on the frontier.

Qwen 3.6-Max-Preview is explicitly labeled a piece in progress. Alibaba stated the mannequin remains to be below lively improvement and expects additional positive factors in future variations. Impartial benchmarking from Synthetic Evaluation places it because the second finest performing mannequin behind Muse Spark—effectively above the median of comparable reasoning fashions in its value tier. The mannequin helps a 256k token context window and handles textual content solely, with no picture enter at launch.