Alibaba’s Qwen team on July 19, 2026 unveiled Qwen 3.8, a 2.4 trillion-parameter multimodal AI model that the company says rivals the world’s leading frontier systems. The team described the model as “second only to Fable 5” — a reference to Anthropic’s top-tier Claude model — though Alibaba has not yet published independent benchmarks to support the claim.
Key points:
- Qwen 3.8 contains 2.4 trillion parameters, making it the second-largest publicly known model behind Moonshot AI’s Kimi K3 at 2.8 trillion.
- Built on a sparse Mixture-of-Experts architecture; first Qwen multimodal model exceeding one trillion parameters. Processes text, images, videos, and documents.
- Preview available through Alibaba’s Token Plan, Qoder coding platform, and QoderWork at 10% of standard pricing.
- Lite plan: 68 for 40,000 weekly credits, six to eight concurrent agents. OpenAI/Anthropic API protocol compatibility claimed.
- Open-weight release promised “soon” without a committed date — would break Alibaba’s pattern of keeping Max-tier models closed-source.
- Active-parameter count at inference not disclosed.
- Predecessor Qwen 3.7-Max scores cited: 92.4% GPQA Diamond, 80.4% SWE-bench Verified, 69.7% Terminal-Bench 2.0 (Qwen 3.8 scores not published).
- Context: Moonshot Kimi K3 (Alibaba holds ~36% stake) and Zhipu GLM 5.2 intensify Chinese frontier competition; Alibaba also noted as Apple Intelligence tech partner in China.
Market note: Alibaba (BABA) closed at 262B market cap). Announcement landed after U.S. market hours on a Saturday.
Source: MLQ News aggregation citing The Decoder, OfficeChai, BuildFastWithAI, and related coverage.