Alibaba Unveils Qwen3.8-Max, Challenging Western AI Giants on Long-Horizon Enterprise Work

Key points

• Qwen3.8-Max is a 2.4-trillion-parameter mixture-of-experts model featuring 95 billion active parameters per processing step ¹².
• Alibaba reports a score of 86.1 on the OSWorld-Verified benchmark, placing it ahead of GPT-5.6 Sol Max and Fable 5 ¹.
• Open weights for the model and a smaller 27B variant are scheduled for release next week, marking a first for Alibaba's Max-class lineup ¹².
• API pricing is set at $2 per million input tokens and $6 per million output tokens, undercutting leading Western proprietary models ¹².

A New Flagship Architecture

Chinese tech giant Alibaba has officially introduced Qwen3.8-Max, its latest flagship multimodal large language model designed explicitly for autonomous software engineering and long-horizon enterprise workflows ¹². Built on a mixture-of-experts (MoE) architecture, the massive system boasts 2.4 trillion total parameters, with 95 billion active parameters engaged during any given processing step ¹². Initially previewed earlier in Shanghai, the model is now commercially accessible via the QwenCloud platform, with open weights slated for release the following week ¹².

This upcoming open-weight release represents a significant strategic milestone for Alibaba. If the weights are distributed under a permissive license, it will mark the first time a model from the high-end Max series becomes available for self-hosted enterprise deployment ¹². Alongside the flagship release, a smaller 27-billion-parameter variant is also planned for open availability ¹². However, industry observers note that the exact licensing terms remain unconfirmed, leaving open the possibility of a restrictive custom license akin to recent releases from other market players ¹.

Autonomous Execution Benchmarks

Alibaba is positioning Qwen3.8-Max not as a casual conversational assistant, but as an autonomous coworker capable of driving multi-day projects ¹². To substantiate these claims, the company released extensive benchmark data highlighting strong performance in agentic computing and complex digital environments ¹². Most notably, Qwen3.8-Max achieves an 86.1 score on the OSWorld-Verified benchmark, edging out prominent competitors like GPT-5.6 Sol Max at 83.2 and Fable 5 at 85.0 ¹.

Beyond computer-use tasks, the model demonstrates notable strength in multimodal processing, document analysis, and professional domains ¹². Company demonstrations highlight ambitious long-horizon feats, including building software repositories from scratch over ten-day spans, reproducing research papers without starter code, and optimizing cryptographic chip designs through hundreds of iterative steps ¹². Nevertheless, third-party analysts caution that while these capabilities point to a powerful new class of persistent agents, independent verification across all benchmark categories is still pending ¹².

Competitive Pricing Pressures

In addition to raw technical metrics, Qwen3.8-Max enters a fiercely contested pricing landscape ¹². The model’s API pricing is established at $2 per million input tokens and $6 per million output tokens, alongside discounted rates for cached inputs ¹². This pricing structure significantly undercuts comparable top-tier Western proprietary offerings, costing less than a third of the combined input-output rates of rival systems ¹².

Such aggressive cost structures are becoming increasingly critical as enterprise AI shifts toward agentic workflows ¹. Because multi-hour autonomous tasks and iterative planning generate millions of tokens per session, inference economics play a decisive role in large-scale deployment ¹². By balancing broad performance profiles with competitive API pricing, Alibaba aims to capture market share among enterprise buyers deploying fleets of autonomous digital coworkers ¹².

Companies mentioned: Alibaba

Primary sources

1. Qwen3.8-Max arrives with a bold claim: it outperforms GPT-5.6 Sol Max and Fable 5 on agentic computer use (venturebeat.com) – Detailing Alibaba's launch of Qwen3.8-Max, its strong performance on agentic benchmarks like OSWorld-Verified, and the strategic implications of its upcoming open-weight release and competitive pricing.
2. Qwen3.8-Max is The Next Chinese Open-Weights Assault on The AI Frontier (trendingtopics.eu) – Providing technical specifications on the model's 2.4-trillion-parameter MoE architecture, specific autonomous task demonstrations, detailed test category comparisons, and tiered API pricing structures.

Powered by News Ranker