Alibaba has unveiled its flagship model Qwen3.8-Max, featuring 2.4 trillion parameters, designed to autonomously handle complex, long-horizon tasks such as reproducing research papers and designing chips. The team plans to release the weights next week.
SynthePulse Insight · AI deep reading
Version 1 · 1 source
Alibaba releases Qwen3.8-Max with 2.4 trillion parameters and 95 billion active, focusing on multi-day autonomous complex tasks. Internal tests show it can autonomously build software, reproduce and surpass papers, and simulate e-commerce profit doubling, but benchmarks are self-reported and independent verification is pending.
On August 3, Alibaba's Qwen team released its flagship model Qwen3.8-Max, with 2.4 trillion total parameters and 95 billion activated per query, based on the Qwen3.5 architecture. The model focuses on independently completing complex tasks over multiple days, rather than single-turn Q&A.
The model was available in preview since mid-July via Token Plan, Qoder, and QoderWork at 10% of the standard price. At that time, the team mentioned the 2.4 trillion parameters and said it ranked second only to Fable 5, but did not release benchmarks. Qwen3.8-Max is the first model in the Qwen-Max series to have public weights.
Weights are planned for release next week on Hugging Face and ModelScope. The model is now available via QwenCloud, supporting the OpenAI Chat Completions format.
The team showcased three autonomous coding cases. In the first, the model spent 16 days building the command-line tool oh-my-cli, turning user requests into GitHub issues, self-assigning, writing code, running tests, and iterating. As of July 30, 2026, it had produced 265 commits, 127 PRs, and 151 issues, with no human intervention.
In the second case, the model received the paper 'Unified Data Selection for LLM Reasoning' without starting code, needing to reproduce and improve results. In about 5 days and 125 hours of compute, it wrote 7,600 lines of code, ran 33 GPU training jobs, reproduced all 6 main results, then tested 18 of its own ideas, surpassing the original method by 2.7 points on the AIME24 math benchmark.
The third case was the WWW2025 multimodal dialogue intent recognition challenge, with 526 human teams. The model fine-tuned multiple Chinese models and Qwen2.5-VL-7B within 24 hours, combined a voting system, and improved accuracy from 0.60 to 0.853 across 45 submissions, surpassing 458 teams.
Two cases tested long-horizon tasks with hundreds of interaction rounds. In the chip design case, the model designed an encryption circuit, optimizing from 8,298 logic gates to 678 after about 500 iterations. After OpenROAD auto-layout, the chip area shrank from 106x106 micrometers to 46x46 micrometers, an 81% reduction. The team said the model still made deep structural modifications after hundreds of iterations.
In the e-commerce simulation case, the model simulated a full year of online retail in E-Commerce-Bench, based on anonymous Taobao and Tmall data. Starting with 100,000 yuan, it had to run multiple stores in parallel, including purchasing, natural language negotiations with suppliers, pricing, returns, and crisis management like typhoons. The supplier pool hid 152 scammers, which the model had to identify. The final balance was 416,252 yuan, 4 times the starting capital, 38% higher than second-place GLM 5.2, and over 2.5 times the predecessor Qwen3.7-Max. The model invested aggressively early in the year, with holiday season net profit exceeding 100,000 yuan.
For multimodality, the model can process documents over 200 pages and videos over 100 hours. The team introduced the RecreationBench benchmark, requiring the model to rebuild applications without source code, only by observing clicks and keyboard interactions, covering Ubuntu, macOS, Windows, Android, and Web. They also released the Qwen-MM-Plugins extension library, adding image/video processing, visual tool use, and multimodal memory to existing agent systems.
The team's published benchmarks show the model approaching or surpassing Claude Opus 4.8, Claude Fable 5, and GPT-5.6 Sol in several categories. PaperBench scored 93, the highest among comparisons, while TerminalBench 2.1 scored 86.6, below GPT-5.6 Sol's 88.8. These are internal runs, with independent verification pending.
The team attributes long-task capability to expanded reinforcement learning training environments, covering multi-day workflows, nested directory structures, and various agent frameworks. The internal score index rose from 0.474 to 0.725, with peak performance at around 4,000 environments, then slightly declining.
Qwen3.8-Max's direct competitor is Kimi K3, released by Moonshot AI on July 27, which is a 2.8-trillion-parameter multimodal MoE model with a 1-million-token context window and open-source weights. Moonshot also released some infrastructure, including attention kernels, MoE communication libraries, and agent tools. Independent tests show K3 significantly lags Western top models in web capabilities and complex math.
Qwen3.8-Max's release intensifies the Chinese open-source model race. Its weights are about to be public, potentially further driving community applications.
This report is based on THE DECODER's retelling; all benchmarks and case results are self-reported by the Alibaba team, and independent verification has not yet been conducted. Some data such as commit counts and accuracy rates come from team statements and have not been third-party confirmed.
Qwen3.8-Max demonstrates the potential of open-source models in long-horizon autonomous tasks, but self-reported benchmarks and cases require independent verification. Its open-source weights will facilitate community evaluation, and the Chinese open-source model race may accelerate.
Primary report
Primary source