gen‑ai.news
← Back
Multimodal

Alibaba Qwen Releases Qwen3.8-Max: A 2.4 Trillion Parameter MoE Model and the Most Capable One in the Qwen Family to Date

Alibaba's Qwen team has officially moved Qwen3.8-Max out of preview and into general availability, bringing with it published per-token pricing and a commitment to release open weights in the coming week. The model is positioned as the most capable in the Qwen family, succeeding the earlier Qwen3 series that already spanned a range of sizes and architectures.

Qwen3.8-Max is built on a mixture-of-experts (MoE) architecture with 2.4 trillion total parameters. MoE designs activate only a subset of parameters for any given token, which allows very large theoretical parameter counts while keeping per-query compute more manageable than a dense model of equivalent size. The approach has become increasingly common among frontier labs looking to scale model capacity without proportional increases in inference cost.

On the input side, the model is multimodal - accepting text, images, and video alongside a 1 million-token context window. That context length places it in a bracket suited to long-document analysis, extended video understanding, and tasks that require holding large amounts of information in a single pass. Exact pricing tiers have been published alongside the general availability announcement, giving developers a clearer picture of what deployment costs will look like before the open weights arrive.

One notable gap at launch is the absence of a formal benchmark table. Independent and third-party evaluations will likely follow once the open weights are publicly released, at which point comparisons against other large MoE models - such as those from Mistral, DeepSeek, and Google - will be easier to make on standardized tasks. Until then, Alibaba's capability claims rest on internal assessments. The open-weight release next week will be a significant moment, as it will allow researchers to run their own evaluations and integrate the model into projects that require local or self-hosted deployment.

Enjoy this story? Get the next one in your inbox.

Twice a week: the most important stories in generative image and video AI, distilled into a 2-minute read.

Free. Unsubscribe any time. No spam, ever.

Your next read

No image
Multimodal

Fine-tune video and image models at scale with NVIDIA NeMo Automodel and 🤗 Diffusers

NVIDIA and Hugging Face have joined forces to bring large-scale fine-tuning of image and video diffusion models into the NeMo Automodel framework, integrated with the Diffusers library. The collaboration aims to make distributed training more accessible for teams working with models that would otherwise be difficult to fine-tune on limited hardware. The result is a more streamlined path from a pretrained diffusion model to a customized one, without requiring deep infrastructure expertise.

No image
Multimodal

Thinking Machines Lab Releases Inkling: A 975B-Parameter Open-Weights Multimodal MoE With 41B Active Parameters And Controllable Thinking Effort

Thinking Machines Lab has released Inkling, a 975B-parameter open-weights multimodal model built on a Mixture-of-Experts architecture that keeps only 41B parameters active at any given time. Licensed under Apache 2.0, it accepts text, image, and audio inputs and offers a 1M-token context window. Rather than competing for top benchmark rankings, the model is positioned as a customizable base with adjustable reasoning depth.

Meta launches Muse Image across its apps and previews Muse Video
Multimodal

Meta launches Muse Image across its apps and previews Muse Video

Meta has rolled out Muse Image, its in-house image generation model, across its family of apps including Instagram Stories, where users can create visuals from text prompts. The company also offered an early preview of Muse Video, signaling its intent to extend the capability to moving images. The moves mark a significant step in Meta's effort to embed generative media tools directly into its social platforms.