XNews

Qwen3.8-Max Release and Upcoming Open-Sourcing

Qwen3.8-Max Release and Upcoming Open-Sourcing

Alibaba Qwen officially released Qwen3.8-Max, a 2.4 trillion-parameter model (95B activated), and announced it and Qwen3.8-27B will be open-sourced next week — the first open-source Qwen Max series release. Demonstrated capabilities include: ran oh-my-cli project autonomously for 16 days; in a 24-hour e‑commerce multimodal recognition contest, after 45 submissions achieved 0.853 accuracy, beating 87% of 458 human teams; in a 365-day simulated e‑commerce environment with seasonality and fraud, achieved 4.16× capital return (over 100k yuan net profit), ranking first; spent 125 hours writing 7,600 lines of code to reproduce a training data selection paper, designed 18 improvements and exceeded the paper’s AIME24 score by 2.71 points; reviewed 1,284 contract clauses in 1 hour; generated an 8‑page high‑fidelity app prototype in one go; autonomously built a data system, wrote factors and ran 6,000 backtests via 330 sub‑agents; processed 200+ page PDFs and 100+ hours of video with graph memory; used vision‑native feedback to self‑correct GUI layouts; and combined coding with GUI operation to replicate apps in black‑box environments. The official comparison omitted recent top Chinese open‑source models; a user supplemented benchmarks.

Key insights

Qwen3.8-Max delivers massive scale with 2.4T total parameters and 95B activated parameters

The model’s architecture contains 2.4 trillion total parameters with 95 billion activated, positioning it among the largest openly‑detailed models. This scale underpins its demonstrated ability to handle extended autonomous tasks and multimodal reasoning.

The model demonstrated fully autonomous programming over 16 days in the oh-my-cli project

Qwen3.8-Max ran an open‑source CLI project completely on its own for 16 days without human intervention, showing long‑cycle autonomous work and self‑evolution capabilities that go beyond typical single‑turn interactions.

In a 24‑hour e‑commerce multimodal intent recognition contest, the model defeated 87% of human participants

The model self‑built and tuned multiple text and vision models, used ensemble voting, and submitted 45 improvements, raising accuracy from 0.60 to 0.853 and outperforming 458 human teams in a real e‑commerce competition.

A 365‑day simulated e‑commerce operation yielded a 4.16× capital return and over 100k yuan net profit

In a realistic e‑commerce simulation with seasonal demand shifts, bargaining, and supplier fraud traps (E‑Commerce Bench), Qwen3.8‑Max autonomously managed operations and achieved the top leaderboard position with a 4.16‑fold return on capital.

The model reproduced a research paper and then exceeded its results, writing 7,600 lines of code and designing 18 improvements

Over approximately 125 hours, Qwen3.8‑Max fully replicated a latest AI training data selection paper from scratch and autonomously iterated 18 improvement schemes, achieving a 2.71‑point higher score than the original paper on the AIME24 math benchmark.

High‑value professional tasks were completed orders‑of‑magnitude faster than human teams

The model reviewed hundreds of contracts and precisely marked 1,284 legal clauses in one hour (versus a typical one‑week human team), generated an 8‑page high‑fidelity app prototype in one pass, and autonomously built a data system, wrote factors, and ran 6,000 backtests via 330 sub‑agents as a quantitative researcher.

Qwen Max series will be open‑sourced for the first time, including Qwen3.8‑Max and Qwen3.8‑27B

Alibaba announced that both Qwen3.8‑Max and a smaller 27B‑parameter variant will be released as open‑source models the week after the release, marking the first open‑source availability for the Qwen Max family.

Official benchmarks omitted recent top Chinese open‑source models, prompting community-driven supplementary comparisons

The release’s comparison tables included only Qwen, Gemini, GPT, and Claude; several competitive Chinese open‑source models were absent. A user manually added those models to the benchmark chart to provide a fuller picture.

Source posts