MiniMax and fal announce livestream on building with H3 video model
The planned session will feature speakers from MiniMax and fal discussing how developers are using H3's multimodal video capabilities, though its date and registration details remain unconfirmed.
By RuntimeWire Staff · Published
Primary source: MiniMax
Why it matters
Fal gives MiniMax a hosted route into developer products without requiring teams to operate H3 themselves. The session could give builders practical guidance on using the model's text, image, video and audio inputs in applications.
Yan Junjie, founder of MiniMax, is promoting a developer session about H3 on fal, the generative-media inference company founded by Burkay Gur and Gorkem Yurtseven. MiniMax's post says the live discussion will examine how creators and developers are building with H3 on fal, including multimodal workflows. The post names Lovis Odin, Ethan Wei and Victor Su-Ortiz as speakers, but the supplied material does not establish the session's calendar date, time or registration link.
Yan founded MiniMax after spending more than six years at SenseTime, where he became a vice president and deputy head of its research institute. His background runs through mathematics and AI research: he earned a doctorate from the Chinese Academy of Sciences' Institute of Automation, completed postdoctoral research at Tsinghua University and, according to MiniMax, has published roughly 200 academic papers. H3 applies that research background to a model designed to handle several parts of audiovisual production through one architecture.
Gur and Yurtseven approached the same market from the infrastructure side. Gur previously worked at Oracle and led machine-learning development at Coinbase, while Yurtseven was a software developer at Amazon. fal handles hosted execution, scaling and API access for models including H3, giving MiniMax another distribution route beyond its own interfaces.
The promotional graphic identifies Odin as a creative engineer at fal, Wei as an AI solutions architect at MiniMax and Su-Ortiz as a GTM engineer at MiniMax. MiniMax says the speakers will break down how creators and developers are using H3 through fal, with an emphasis on multimodal inputs.
How fal distributes H3
MiniMax announced H3 on July 31, 2026, positioning it as a general-purpose video model that can interpret text, images, video and audio within a unified context. MiniMax says H3 can generate clips of up to 15 seconds with native stereo audio and 2K output. Its intended uses range from text-to-video generation to motion transfer, targeted editing and reference-driven production.
MiniMax has expanded H3's distribution through local workflows, third-party creative products and hosted inference services such as fal. The materials do not establish whether fal's hosting arrangement is exclusive or represents a newly negotiated partnership.
fal describes itself as a Day 0 H3 ecosystem partner. Its H3 landing page lists text-to-video, image-to-video and reference-to-video as distinct hosted options. The service gives developers a way to combine text, images, video and audio when building applications around H3.
Why fal fits H3's distribution strategy
Open weights can broaden distribution, while hosted inference removes operational work that prevents many teams from using large video models. Running a model locally can appeal to researchers and developers who need control over the stack. A serverless API is usually the faster route for a product team that wants to ship a feature without procuring GPUs, maintaining inference software or managing traffic spikes.
fal disclosed $23 million across its seed and Series A rounds in 2024. fal later announced a $140 million Series D in December 2025, led by Sequoia with participation from Kleiner Perkins and NVIDIA, and said at the time that its team had grown to 70 employees. The funding gave fal more capacity to compete as an inference provider for generative-media applications, where access to a new model can carry similar commercial value to ownership of the model itself.
For MiniMax, fal offers a route into developer products that may never use MiniMax's own consumer interfaces. For fal, H3 adds a model with native audio and broad reference handling to a catalog where developers can compare competing video systems through similar APIs. Both companies stand to gain when experimentation becomes paid inference traffic.
What the session needs to show
H3's broad specification creates a demanding test. Fal's H3 page says image references can preserve identity, video references can supply motion, audio references can support voice transfer or cloning, and natural-language instructions can direct edits. A useful developer session would show how reliably those controls survive real API calls, how references are ordered and cited in prompts, and how much iteration is required before a clip is usable.
Independent evaluation still trails H3's distribution. No independent benchmark validating MiniMax's cost or quality claims was located in the reviewed materials. MiniMax has instead focused the rollout on practical access through local workflows, third-party creative products and hosted services such as fal.
That sequence reflects H3's distribution strategy. Video models will be judged inside editing pipelines, applications and commercial workloads, where predictable control and generation cost carry more weight than a curated launch clip. The livestream gives MiniMax and fal a venue to explain how developers can turn H3's capabilities into working software. The supplied post does not establish the session's calendar date, time, registration link or customer examples.