From Closed Giants to Practical, Open Multimodal AI
Multimodal AI generation has quickly moved from research showcase to something most product teams can realistically deploy. Two recent launches capture that shift. ByteDance’s Lance offers a 3 billion active parameter model that can understand and generate images and video, while Google’s Gemini Omni Flash and Gemini 3.5 Flash emphasize speed, agentic behavior, and rich media creation. Until now, many of the most capable video generation AI systems have been locked behind restrictive licenses, proprietary APIs, or heavyweight infrastructure requirements. That kept smaller companies from experimenting, even as demand for visual search, AI filmmaking, and automated editing surged. The new wave of open source AI models and efficient AI models changes the equation: Lance’s Apache 2.0 license and Google’s faster, cheaper-to-run Flash variants both aim to make multimodal AI a daily tool, not just a demo on a conference stage.
Lance: An Apache 2.0 Multimodal Model Built for Builders
Lance is ByteDance’s bid to make multimodal AI feel genuinely usable for teams that care about control and integration. At 3 billion active parameters, it targets a middle ground—large enough to handle image understanding, video understanding, image generation, image editing, video generation, and video editing, but small enough to test and ship without turning every roadmap discussion into a GPU capacity review. Trained from scratch with a staged multi-task recipe on no more than 128 A100 GPUs, Lance uses a shared multimodal sequence for text, images, and video, plus separate experts for understanding and generation. The technical design supports one continuous creative workflow: ingest visuals, generate new content, and edit existing assets in the same system. Crucially, the Apache 2.0 license and downloadable checkpoints let startups embed Lance directly into products, fine-tune for niche styles, or keep visual understanding close to proprietary customer data.
Gemini Omni Flash and 3.5 Flash: Speeding Up Multimodal Agents
Google is pushing in a different but complementary direction: ultra-fast, highly capable multimodal AI inside polished products. Gemini 3.5 Flash, now the default model in the Gemini app and in AI Mode in Google Search, is designed as an efficient AI model that rivals larger flagships on multiple dimensions while excelling at coding, agentic tasks, and multimodal understanding. Gemini Omni, meanwhile, focuses squarely on video generation AI. Its first variant, Gemini Omni Flash, can take combinations of images, audio, video, and text to create realistic, knowledge-grounded videos that can be edited conversationally over multiple turns. Integrated into Google Flow—an AI creative studio that already combines Veo, Imagen, and Gemini—as well as YouTube Shorts and YouTube Create, Omni Flash aims to feel like a creative partner: understanding scenes, applying concepts like gravity and fluid dynamics, and maintaining context as users refine each shot.
Why Open, Efficient Multimodal Models Matter for Developers
For developers, the real story is how these releases lower the barrier to building sophisticated multimodal AI generation into products. Lance’s open source AI model under Apache 2.0 removes licensing uncertainty that often stalls commercial experiments, especially in areas like ad creation, visual search, and editing workflows. Teams can host, modify, and redistribute the model, gaining fine-grained control over behavior and data handling. On the proprietary side, Gemini Omni Flash and Gemini 3.5 Flash show what happens when efficient AI models are tightly integrated into agentic interfaces: tools like Google Flow shift from single-prompt playgrounds to persistent workspaces that remember project state, orchestrate assets, and help users progress across multiple steps. Together, these trends signal a more competitive landscape. Closed providers must compete not only on raw performance, but also on speed, economics, and how well their tools plug into real creative and business workflows.
The Next Phase: Multimodal AI on Everyday Hardware
The common thread across Lance and Google’s Flash models is efficiency. Lance’s 3B scale and training budget indicate a deliberate push away from frontier-sized models toward something easier to deploy on consumer-grade or modest cloud hardware. Google’s use of the Flash branding for both Gemini 3.5 Flash and Gemini Omni Flash emphasizes faster inference and lower perceived latency, which becomes crucial as agentic software chains many small model calls together. This matters because multimodal AI only becomes truly transformative when it is cheap and responsive enough to sit behind everyday tools—marketing dashboards, video editors, design platforms—without imposing noticeable friction. Open models like Lance let companies bring that capability in-house, while hosted offerings like Gemini give instant access to advanced video generation AI. The result is a broader, more practical ecosystem where multimodal AI can be embedded wherever there is text, image, or video to understand or create.