Alibaba claims Qwen3.8 is second only to Claude 3.5, but offers no benchmarks to back it up

Alibaba announced Qwen3.8, claiming it ranks among the world's most advanced large language models, specifically positioning itself as a close second to Anthropic's Claude 3.5. Yet the company released no technical documentation, performance benchmarks, or other substantive evidence to support the assertion.

The timing and vagueness drew immediate scrutiny, arriving just days after Alibaba-backed rival Moonshot launched Kimi K3 with detailed architectural breakdowns, comprehensive benchmark comparisons, transparent pricing, and a confirmed July 27 date for open-weight release. By contrast, Alibaba's announcement consisted largely of an X post containing grammatical errors—including what appeared to be conflating "compatible" with "comparable"—and a loose promise that weights would become available "soon."

The benchmarking gap and strategic timing#

Industry observer Julien Simon, an AI partner at Fortino Capital and former Hugging Face evangelist, framed the core issue as a "checkpoint gap": the interval between a frontier claim and the actual evidence that can verify it. Kimi K3 narrowed that gap with discipline—published benchmarks, pricing, and a fixed release date. Qwen3.8 left the gap wide open with no numbers, no timeline, and no verifiable claims.

"Every claim in that launch can be graded against a calendar," Simon wrote in The AI Realist newsletter of Kimi K3's approach, contrasting it with Qwen3.8's deliberately vague "open-weight soon" language. An undated gap, he argued, functions as a vendor option—allowing Alibaba to collect the open-weight narrative while postponing the moment when real performance data could be checked.

The stakes are significant: Alibaba was Moonshot's largest investor as of 2024, having invested roughly $800 million for a 36% stake. Kimi K3's launch reportedly triggered over $3 trillion in losses across global semiconductor stocks within days. Now Alibaba is announcing a marginally smaller model (2.4 trillion versus 2.8 trillion parameters) with a marginally better ranking claim, using identical open-weight language—but without the verifiable details.

Why the ambiguity may be intentional#

Simon suggested that Alibaba faced a strategic bind: publishing real benchmarks could either make the in-house model look superior to its own major investment (damaging Moonshot's profile during IPO planning) or reveal it falls short of Kimi K3 (undermining its headline claim). Withholding numbers sidesteps both outcomes. The announcement, in this reading, was less about showcasing the model and more about inserting Alibaba into headlines dominated by a competitor.