Alibaba Announces 2.4T Open-Weight Model Qwen3.8

On July 19, Alibaba's Qwen team officially previewed Qwen3.8, which will be released as an open-weight model. With approximately 2.4T parameters, it is the largest model Qwen has ever trained. The company claims its capabilities rival frontier models and rank just below Fable 5 in internal evaluations, with the model still evolving daily. This announcement has drawn significant attention, with multiple authors agreeing that it marks Alibaba's aggressive return to the open-source AI race with a massive parameter scale.

Key Details and Benchmarking Context

According to the official preview, Qwen3.8 has about 2.4T parameters. @teortaxesTex noted that Qwen 3.8-Max shows significant improvements in coding and cowork capabilities over Qwen 3.7-Max, and that 2.4T is the largest scale among publicly known models in China. They added that Qwen3.7-Max was previously the highest-ranking Chinese model on the ECI leaderboard, predicting Qwen 3.8 Preview could reach an ECI score of 156. @aigclink provided benchmarking context: tests were based on 400 real tasks from Alibaba's agentic products using isolated ECS and multiple cross-validations. @ugcfast pointed out that the model is multimodal, and Alibaba claimed its pricing is about one-tenth of Fable 5's. However, in a July 20 post, @bindureddy cited the scale as "about 3T"—near Sol/Opus level—and predicted the open-source community would find more efficient deployment methods within 12–15 weeks, though short-term running costs would remain high. This is the only differing parameter figure in this cluster.

Preview Version and Access Channels

Qwen3.8-Max-Preview is already available on Alibaba Token Plan, Qoder, and QoderWork. @zephyrz9 mentioned that the preview is open to Qwen Token Plan subscribers in mainland China. @aigclink also shared practical results of using Cline with the Qwen3.8 preview to build an "AI employee" project: the backend and frontend both succeeded on the first try with only minor modifications.

Background and Impact

@智东西 contextualized Qwen3.8 within the recent wave of dense updates from Chinese LLMs. Authors like @measureplan and @aigclink unanimously view this as a definitive signal of Alibaba re-entering direct open-source competition.

2026-07-19 ~ 2026-07-21 · 22 related posts

Full story(17 episodes)→

Primary sources

8 near-duplicate retellings: haider1 · Alibaba_Qwen · kimmonismus · vista8 · MaziyarPanahi · bodonoghue85 · measure_plan · baseten