en.Wedoany.com Reported - Alibaba on Monday released its largest AI model to date, Qwen3.8-Max, expanding its enterprise AI product portfolio. The open-weight model adopts a Mixture-of-Experts (MoE) architecture with 2.4 trillion total parameters, activating only approximately 95 billion parameters during inference. It is designed to support coding, reasoning, and multimodal tasks with higher inference efficiency, targeting software engineering and knowledge-intensive business workloads.
As planned, the open-weight version will be released next week via Alibaba Cloud's Model Studio. Alibaba stated on X that the model is "one of the most powerful models today, comparable to leading frontier AI models, second only to Claude Fable 5."
In terms of benchmark testing, Alibaba compared Qwen3.8-Max against Anthropic's Claude Opus 4.8, Claude Fable 5, and OpenAI's GPT-5.6 Sol on coding benchmarks, including SWE-bench Pro and Alibaba's custom evaluation benchmark NL2Repo-Bench. For evaluating competitors, each vendor's own coding tools were used—Claude Code for Anthropic models and Codex for GPT-5.6 Sol—and the highest published scores for each competitor under available configurations were reported.
Charlie Dai, Vice President and Principal Analyst at Forrester Research, believes this release demonstrates that Alibaba is narrowing the gap with proprietary model leaders, but what deserves more attention is the rapid maturation of open-weight models. He noted that in software engineering, domain customization, sovereignty, and cost-sensitive deployment scenarios, enterprises now have reliable alternatives to frontier proprietary models, where openness often matters as much as absolute model performance.
Alibaba also reported results from three unsupervised multi-day coding project tests: the model started from blank project folders and completed tasks independently with no human assistance throughout the process. One of these projects reportedly took 16 days to complete autonomously. Enterprise application directions include legal compliance, financial analysis, engineering design, quantitative research, and multimodal content creation. Alibaba stated that the model is designed to complete entire business workflows rather than individual AI-assisted tasks.
Amit Jena, AI Development Manager at Kanerika, questioned the claim of "autonomously completing a software engineering project in 16 days": "What was done in those 16 days? How many human interventions occurred? Was the output reviewed through code review?" He also cautioned that open weights and open API endpoints are two different things—"without a code repository, license, and model card, open weights are merely an intention."
Regarding the design of activating only approximately 95 billion parameters, Dai believes that for enterprise buyers, inference efficiency has become more important than raw model size, as it can significantly reduce service costs and infrastructure requirements. Jena, however, stated that efficiency is no longer the key issue—evaluating throughput is the real constraint. Nitish Tyagi, Senior Principal Analyst at Gartner, pointed out that this release indicates competitive pressure on AI deployment costs. Gartner previously predicted that without stronger cost controls, AI coding expenses could exceed the salaries of average developers. Tyagi believes that the combination of open weights, Mixture-of-Experts architecture, and a 1-million-token context window makes AI-assisted software development more economically viable.
Tyagi also cautioned that enterprises evaluating production deployment cannot consider only inference costs. Many organizations outside China may be cautious about models hosted in China, and deployment through hyperscale cloud providers or local infrastructure incurs additional costs. Open-weight models typically lack the indemnification protection offered by commercial AI vendors, requiring enterprises to establish their own security, governance, and code-scanning controls to identify copyright and intellectual property risks before production deployment.
Jena also noted that Qwen3.8-27B, released alongside the flagship model but receiving little media attention, may be a more practical choice for most organizations, as it can run on their own infrastructure and be fine-tuned with their own data. Dai suggested that enterprise leaders prioritize evaluating transparency and total cost of ownership over headline numbers. The core question is whether Qwen3.8 can deliver measurable business outcomes, enterprise-grade reliability, lower total cost of ownership, and digital sovereignty options compared to competing models.
