Qwen3.8 Goes Live With 2.4T Parameters - Alibaba Claims Second to Fable 5, Open Weights Coming

Alibaba's Qwen team announced Qwen3.8 today via the official @Alibaba_Qwen X account, making a 2.4-trillion-parameter multimodal model available in limited developer preview. Qwen3.8 is live now on Token Plan, Qoder, and QoderWork at 10% of standard pricing, with open weights promised for an unspecified future date. No independent benchmarks have been published. Alibaba places Qwen3.8 second globally - trailing only Claude Fable 5 in overall capability - based entirely on its own internal evaluations.

Preview Pricing Starts at $6 - OpenAI and Anthropic Protocols Both Work

Token Plan entry costs $6 per month for 2,500 credits every seven days, or $68 per month for 40,000 credits at the Pro tier with support for six to eight concurrent agents. All three access points share the same 10% preview rate. Alibaba's Qwen models already power Apple Intelligence in China, and Qwen3.8-Max-Preview ships with OpenAI and Anthropic API protocol support, meaning teams running existing agentic workflows can point them at Qwen3.8 without rewriting their integration code or switching providers.

Qwen3.8 Is Alibaba's First Multimodal Model Above 1 Trillion Parameters

Qwen developer Shuai Bai confirmed Qwen3.8 as the Qwen team's first multimodal model with over one trillion parameters, capable of processing images, video, and documents alongside text. Alibaba says Qwen3.8 outperforms Qwen 3.7-Max on coding and productivity tasks including full-stack development, data analysis, and office document automation - but no benchmark numbers accompany that comparison. No activated parameter count has been published. Previous Qwen previews followed the same limited-access pattern before weights arrived: Qwen3-Max-Preview launched in September 2025 as the first Qwen model above 1T, and Qwen 3.6 Max Preview appeared in April 2026 before broader release.

Alibaba Claims Second to Fable 5 - No Model Card, No Published Benchmarks

Qwen's official announcement described Qwen3.8 as "second only to Fable 5" - the same claim Moonshot made when announcing Kimi K3 two days earlier. No model card, no report, no weights. Outlets like Artificial Analysis and LMArena need to run independent tests before anyone outside Alibaba can verify that ranking - and no timeline for third-party evaluation has been set. Kimi K3's version of the same claim held up reasonably well when put to a direct benchmark comparison against Fable 5; Qwen3.8 has not had that test yet.

Kimi K3 Open Weights Land July 27 - Qwen3.8 Has No Date

Open weights for Qwen3.8 arrive "soon" - no specific date, no license terms announced. Kimi K3, which launched two days before Qwen3.8, set a firm July 27 date for its open-weight release. Qwen3.8 has made no comparable commitment. Alibaba owns approximately 36% of Moonshot AI, the lab behind Kimi K3, which turns what looks like standard market competition into something closer to a portfolio rivalry - two Alibaba-backed labs both racing for the title of best open Chinese frontier model in the same week, a dynamic Moonshot's founders probably did not anticipate when they took Alibaba's capital.

GLM 5.2 from Z.AI made its open-weight push in June. Kimi K3 followed in July. Qwen3.8 joins the queue now, making three major Chinese labs in roughly three months each claiming a spot near the top of the global leaderboard and promising weights that bring their models within reach of developers who prefer self-hosted deployment over API costs. Whether Qwen3.8's performance matches its internal billing is a question independent evaluators will answer - open weights or not.