Alibaba's Frontier AI
Alibaba Unveils Multimodal Qwen 3.8 AI, Reaching 2.4 Trillion Parameters and Rivaling Frontier Systems
Anonymous
Key Points
1
Moonshot AI, which Alibaba holds a 36% stake in, released the Kimi K3 model at 2.8 trillion parameters earlier this year and briefly held the title of the largest open-source model.
2
Qwen 3.8 marks the team’s first multimodal model to exceed one trillion parameters, underscoring a milestone for scale in its multimodal lineup.
3
Qwen 3.8 is described as the second-largest publicly disclosed model, behind Anthropic’s Fable 5, highlighting its prominence in the open disclosures of large models.
4
The Qwen 3.8-Max-Preview was announced during WAIC in Shanghai, situating the rollout within China’s major AI summit context.
5
Preview pricing details include access at 10% of standard pricing via Token Plan/Qoder/QoderWork, with Lite and Pro tiers priced at $6 for 2,500 credits/week and $68 for 40,000 weekly credits (supporting six to eight concurrent agents).
Alibaba has launched Qwen 3.8, a massive 2.4-trillion-parameter AI model that the company says rivals the world's best AI systems. South China Morning Post reported the model's preview — called Qwen 3.8-Max-Preview — claims to trail only Anthropic's Fable 5 in overall performance.
The announcement came at WAIC, China's major AI summit in Shanghai. It marks a significant moment in the global race to build ever-larger AI models, with Chinese companies now openly competing with American labs at the frontier.
A 2.4 Trillion-Parameter Model Built on Sparse MoE
Qwen 3.8 is built on a Mixture-of-Experts (MoE) architecture. In plain terms, MoE means the model is split into many smaller "expert" networks, and only some of them activate for any given task. This makes very large models cheaper to run. According to MLQ.ai, Alibaba has not yet disclosed how many parameters are active at inference — a key detail that shapes real-world performance and cost.
The model is multimodal, meaning it can handle text, images, videos, and documents in one system. MLQ.ai noted this is the Qwen team's first multimodal model to break one trillion parameters, a milestone for the lineup. The 2.4 trillion total figure makes it the second-largest publicly disclosed model in the world.
Moonshot AI's Kimi K3 Adds Heat to the Chinese AI Race
Alibaba is not alone in pushing scale. Moonshot AI, in which Alibaba holds a 36% stake, released its Kimi K3 model earlier this year at 2.8 trillion parameters. Yahoo Finance reported that Kimi K3 briefly held the title of the largest open-source model in the world.
Both Alibaba and Moonshot are moving to widen developer access. Kimi K3 is planned as an open-weight release, meaning developers can download and run it themselves. This mirrors a broader strategy among Chinese AI labs to attract developers away from closed Western systems.
Preview Pricing Tiers Target Developers and Enterprises
Alibaba is offering early access to Qwen 3.8-Max-Preview at 10% of standard pricing through its Token Plan, Qoder, and QoderWork platforms. South China Morning Post cited two tiers: a Lite plan at $6 for 2,500 credits per week, and a Pro plan at $68 for 40,000 weekly credits. The Pro tier supports six to eight concurrent AI agents running at the same time.
The discounted preview pricing signals Alibaba wants fast developer adoption before a full commercial launch. Locking in developers early is a common strategy in the AI platform wars, where switching costs rise once teams build products on a specific model.
Key Benchmarks Still Missing as Investors Watch Closely
Alibaba has not released full benchmark results for Qwen 3.8. That matters because benchmark scores are the main way the industry compares models. Without them, the claim of "near-frontier" performance rests largely on Alibaba's own word. The missing active-parameter count for the MoE design adds another gap in the public picture.
Investors are watching how Alibaba's AI stack holds up against both American rivals and domestic Chinese competition. Yahoo Finance noted that the open-weight strategy for both Qwen 3.8 and Kimi K3 could speed enterprise adoption — but also raises questions about how Alibaba will monetize models that anyone can download and run for free.
Sentiment
Mostly PositiveNEGATIVENEUTRALPOSITIVE
Publishers
13
Articles
12
Reach
25
A
AnonymousAuthor