Alibaba’s small AI model matches OpenAI’s GPT-5.6 Luna

Alibaba ↗ 0.73% has released a 27-billion-parameter AI model that performs on par with OpenAI's GPT-5.6 Luna and nearly matches trillion-parameter systems from DeepSeek and Zhipu, according to benchmark data published on Monday — the latest example of Chinese labs closing the gap with American rivals using far fewer resources.

Small Model, Big Results

The Qwen3.8-27B, whose model weights were released last Friday, scored 52 on the Artificial Analysis Intelligence Index, matching GPT-5.6 Luna, which OpenAI has billed as the most cost-efficient model in its latest flagship series. It nearly matched DeepSeek-V4-Pro-0813, released last week with 1.7 trillion parameters, and Zhipu's 753-billion-parameter GLM-5.2, launched in June.scmp+2

On Artificial Analysis' Agentic Index, which measures performance in AI agent-focused workflows, the small Alibaba model scored 51 — outperforming OpenAI's mid-tier GPT-5.6 Terra and Anthropic's Claude Opus 4.8, released in May.gigazine+1

The model can run on consumer-grade hardware, a point that has drawn attention from developers interested in local AI deployment. Its API pricing through Alibaba's cloud is effectively free, at zero dollars per million input and output tokens.artificialanalysis

A Broader Pattern of Chinese Competitiveness

The Qwen3.8-27B result is part of a wider trend that has unsettled Silicon Valley. Bloomberg reported in April that Chinese AI models are now "cheaper and more adaptable than the preeminent US platforms" and that studies suggest they are "almost as proficient". Chinese developers have focused on systems that perform near the frontier without requiring the most powerful hardware, and have bet heavily on open-weight distribution to accelerate adoption.bloomberg

The pattern has intensified this summer. Alibaba's larger Qwen3.8-Max, a 2.4-trillion-parameter model unveiled on August 3, claimed to outperform GPT-5.6 Sol Max on autonomous software engineering benchmarks. Meanwhile, Moonshot AI's Kimi K3 model, launched in early August with 2.8 trillion parameters, overwhelmed its own infrastructure with demand within two days.facebook+2

What It Means for the AI Race

The competitive dynamic challenges a US industry model that has relied on massive capital expenditure and proprietary systems. OpenAI does not disclose the parameter counts of its models, making direct architectural comparisons difficult. But the fact that a 27-billion-parameter model — small enough to run locally — can trade blows with frontier commercial systems raises questions about the returns on hundreds of billions in infrastructure spending by American tech companies.scmp