Qwen 3.8: 2.4 Trillion Parameters and a Promise to Open the Weights
Alibaba's Qwen team unveiled Qwen 3.8 on July 19: a 2.4 trillion parameter multimodal model, the second-largest publicly known model after Moonshot's 2.8T Kimi K3, which shipped just days earlier. The claim doing the rounds: Alibaba says it is second only to Fable 5. No third-party benchmarks exist yet, so that claim currently rests entirely on internal evals. Worth repeating until the weights drop.
Two things make this more than another giant-model press release. First, open weights are promised soon, which would break Alibaba's long-standing pattern of keeping its Max-tier models closed. If it lands, the open-weight frontier becomes a two-horse race between 2.4T and 2.8T models within a single month. Second, this is Qwen's first multimodal model above 1T parameters, taking images, video and documents, and it is aimed squarely at coding, full-stack development, data analysis and office workflows. That is agent bread-and-butter, not chatbot territory.
The timing is a direct shot at Moonshot. Kimi K3 is so oversubscribed that Moonshot suspended new subscriptions this weekend because demand outran inference capacity. Every frontier lab right now has the same problem: the models are ready, the GPUs are not. Qwen 3.8 is in preview via Alibaba's Token Plan, Qoder and QoderWork at 10% of standard pricing. The Hacker News thread sat at 706 points.
Announcement: https://x.com/alibaba_qwen
Coverage: https://the-decoder.com/alibabas-qwen-takes-on-kimi-k3-with-open-weight-qwen-3-8-says-model-is-second-only-to-fable-5/
← Back to all articles
Two things make this more than another giant-model press release. First, open weights are promised soon, which would break Alibaba's long-standing pattern of keeping its Max-tier models closed. If it lands, the open-weight frontier becomes a two-horse race between 2.4T and 2.8T models within a single month. Second, this is Qwen's first multimodal model above 1T parameters, taking images, video and documents, and it is aimed squarely at coding, full-stack development, data analysis and office workflows. That is agent bread-and-butter, not chatbot territory.
The timing is a direct shot at Moonshot. Kimi K3 is so oversubscribed that Moonshot suspended new subscriptions this weekend because demand outran inference capacity. Every frontier lab right now has the same problem: the models are ready, the GPUs are not. Qwen 3.8 is in preview via Alibaba's Token Plan, Qoder and QoderWork at 10% of standard pricing. The Hacker News thread sat at 706 points.
Announcement: https://x.com/alibaba_qwen
Coverage: https://the-decoder.com/alibabas-qwen-takes-on-kimi-k3-with-open-weight-qwen-3-8-says-model-is-second-only-to-fable-5/
Comments