Quick Facts
- Alibaba’s Qwen3.8-Max-Preview uses 2.4 trillion parameters and a sparse Mixture-of-Experts architecture, making it the company’s first multimodal model above 1 trillion parameters.
- The model is available now through Alibaba’s Token Plan subscriptions, priced from $6 to $68 per month, at 10% of standard pricing during the preview period.
- Alibaba has published no benchmark data to support its claim that Qwen3.8 ranks second only to Anthropic’s Claude Fable 5.
Alibaba Group unveiled a preview of Qwen3.8-Max at the World Artificial Intelligence Conference in Shanghai on Sunday, claiming the model is second only to Anthropic’s Claude Fable 5 among frontier AI systems. The company offered no benchmark data, model card, or independent test results to support that ranking.
The model carries 2.4 trillion parameters and runs on a sparse Mixture-of-Experts architecture. It supports a 1-million-token context window, thinking mode, function calling, and built-in tools, and can process images, video, and documents alongside text.
Developer Shuai Bai called Qwen3.8 “the team’s first multimodal model with more than 1 trillion parameters.” Alibaba says the model should outperform its predecessor, Qwen3.7-Max, on coding and productivity tasks including full-stack development, data analysis, and office workflows.
Benchmark Claims Without Evidence
Alibaba’s self-assessment is a departure from its approach in May, when it published a fuller set of performance results alongside Qwen3.7-Max. No outside leaderboard has scored Qwen3.8. The most recent Qwen model to receive an independent ranking, Qwen3.7-Max, sits far from the top of LMArena’s crowd-voted list, where Claude Fable 5 currently holds first place.
On SWE-bench Verified, Fable 5 holds a score near 95%, leaving a gap of roughly 15 points over the next closest competitors. Chinese rivals Kimi K3, Inkling, and MiniMax M3 all released models this month without closing that margin. Alibaba is asking buyers to accept a leap from the middle of the pack to second in the world based on its own word.
Alibaba also has not disclosed how many of the 2.4 trillion parameters are active at inference time, a critical figure for sparse MoE models where only a fraction of parameters engage per query.
Priced for Broad Adoption
Qwen3.8-Max-Preview is available now through Alibaba’s Token Plan subscription service, its Qoder coding platform, and the QoderWork productivity suite at 10% of standard pricing. Token Plan tiers range from a $6 Lite plan offering 2,500 credits every seven days to a $68 Pro tier providing 40,000 credits per week with support for six to eight concurrent agents. Open weights are coming soon, though Alibaba has not set a date.
Shares of Alibaba rose as much as 5.4% in Hong Kong on Monday, moving from $114.97 to $120.84 after the announcement. Investor sentiment also reflected a reported partnership with Apple that could integrate Qwen into consumer devices.
China’s Model Race Heats Up
The Qwen3.8 preview arrived three days after Beijing Moonshot AI released Kimi K3, a 2.8-trillion-parameter model that recently pushed Claude Fable 5 into second place on a prominent AI coding leaderboard. Alibaba is among Moonshot’s backers, making the rivalry partly internal.
Moonshot disclosed its architecture, API pricing, evaluation conditions, and a weight release deadline before Alibaba made comparable disclosures for Qwen3.8. Bank of America analysts led by Alex Liu wrote that Kimi K3 has raised the capability ceiling for Chinese models and that Alibaba’s position as China’s open-source leader may face fresh tests.
Alibaba CEO Eddie Wu said the company’s Qwen consumer interface now exceeds 300 million monthly active users. Both Qwen3.8 and Kimi K3 mark the arrival of Chinese AI labs in the multi-trillion-parameter tier, a scale previously associated mainly with the largest U.S. research labs.
Read more: Alibaba previews Qwen3.8, claims it’s second only to Claude Fable 5
This article was written by an AI agent. Spotted an error? Send a correction and we will fix it.