// AI NEWS · AI VIDEO

HeyGen Brings Its 18B-Parameter Avatar IV Model to Google Cloud TPUs, Cutting the Cost per Minute of Avatar Video

In an August 13, 2026 Google Cloud engineering write-up, HeyGen detailed how it ported Avatar IV — its 18-billion-parameter image-to-video avatar model — to eight-chip Trillium (v6e) TPUs, hitting a 1.86x speedup over its first working version and up to 25% lower cost per generated minute versus an 8×H100 setup.

2026-08-14 · by Moe Ameen

What happened

On August 13, 2026, the Google Developers Blog published a joint engineering write-up with HeyGen describing how the avatar-video company moved Avatar IV — its image-to-video model — onto Google Cloud's Trillium (v6e) TPUs. Avatar IV is the model that turns a single photo and an audio track into a talking, moving person; the post is a rare look under the hood at what it takes to serve it at scale.

The technical details are specific. Avatar IV runs on more than 18 billion parameters across three components: a diffusion transformer that renders motion, a second transformer for super-resolution, and a VAE decoder that produces the final pixels. The two transformers alone exceed 36 GB of bf16 weights, more than the 32 GB of high-bandwidth memory on a single Trillium chip, so HeyGen sharded the weights across an eight-chip host using FSDP and added Ulysses sequence parallelism. The port ran through torchax — described as a PyTorch frontend on JAX — so HeyGen's production model code runs unmodified, dispatched onto JAX arrays and compiled by the XLA compiler. Output is 720p or 1080p at 25 frames per second, streamed in chunks as they finish so playback can start before the full clip renders.

The payoff is speed and cost. Working with Google Cloud's AI infrastructure performance team, HeyGen tuned custom kernels and the XLA compiler to reach a 1.86x speedup over its first working TPU version. The company reports the Trillium setup performs comparably to its 8×H100 production configuration while being up to 25% more cost-efficient per minute of generated video. Treat the exact figures as vendor-reported engineering numbers; the direction — cheaper, faster avatar minutes — is the point.

Avatar IV itself is not new. HeyGen introduced it in 2025 as its most advanced image-to-video model, adding authentic hand gestures, voice-synced emotion, and support for tilted heads, profiles, and angled source photos, and later opened it up via an Avatar IV API. What changed on August 13 is the economics of running it: the same model, served more cheaply. HeyGen said in June 2026 that it had crossed $200 million in annual recurring revenue, and infrastructure cost is exactly the lever that keeps a per-minute generation business profitable as volume grows.

Why it matters for creators

  • Cheaper avatar minutes eventually reach the price you pay. HeyGen serving Avatar IV up to 25% cheaper per minute is the kind of infrastructure win that shows up later as more generous credit allowances or lower per-minute costs — good news if avatar video is part of your cadence.
  • It signals Avatar IV is a long-term bet, not a stopgap. Companies do not do deep TPU ports of models they plan to retire; the effort says HeyGen intends to keep Avatar IV — single-photo, image-to-video — central for a while.
  • The realism bar for a talking photo keeps rising. Hand gestures, voice-synced emotion, and angled-photo support mean an avatar built from one still keeps getting harder to distinguish from filmed footage — which raises both the creative ceiling and the disclosure stakes.
  • It is still just the render. Faster, cheaper generation produces a raw MP4 more efficiently; it does nothing about captioning it, reframing it for each feed, or publishing it on a schedule. The generation step got cheaper; the distribution step did not move.
  • Infrastructure cost, not model quality, is becoming the competitive front. When the leading avatar models are all "good enough," the fight shifts to who can serve a minute of video cheapest — a race that favors whoever turns raw generation into finished, published content most efficiently.

How to act on this with Kompozy

The headline number here is a cost-per-minute win, and that framing is worth borrowing — because for a creator, the cost that actually hurts is not the render, it is everything you still do by hand after it. A cheaper, faster Avatar IV minute lands you at exactly the same place a more expensive one did: a bare talking-head MP4 with no captions, no per-feed reframe, and no way onto your platforms. HeyGen just optimized the one step that was never your bottleneck. [Kompozy](/) is built for the steps that are. It runs Avatar-IV-class avatar generation natively inside [Persona Shorts](/glossary/persona-shorts) and its longer-form Persona HeyGen format, then auto-captions the take with brand-exact styling, reframes it to 9:16, 1:1, and 16:9, and — through [Persona Frames](/glossary/persona-frames) — composites the avatar as a movable layer inside a pixel-exact [HyperFrames](/glossary/hyperframes) template.

That is where the real minutes get saved. From one [Persona Brief](/glossary/persona-brief), the same idea that became an avatar clip also becomes a carousel, a quote card, an X thread, a [blog article](/glossary/output-buckets), and an email newsletter, all in one voice, then schedules and publishes across the eight social platforms plus blog and email on [Autopilot](/glossary/autopilot). Let HeyGen and Google fight over the price of a rendered second; the leverage for a creator is turning that second into a week of on-brand, multi-platform content without touching six separate apps.

Quick takeaways

  • On August 13, 2026, HeyGen and Google Cloud published a write-up on porting Avatar IV to Trillium (v6e) TPUs.
  • Avatar IV is an 18B+ parameter image-to-video model: a diffusion transformer for motion, a super-resolution transformer, and a VAE decoder, outputting 720p or 1080p at 25 fps.
  • The two transformers exceed 36 GB of bf16 weights, so HeyGen FSDP-sharded them across an eight-chip host and used Ulysses sequence parallelism.
  • The port used torchax (a PyTorch frontend on JAX) so production model code runs unmodified under the XLA compiler.
  • HeyGen reports a 1.86x speedup over its first working TPU version and up to 25% lower cost per minute versus an 8×H100 setup.
  • Avatar IV turns a single photo plus audio into a talking video with hand gestures and voice-synced emotion; it launched in 2025 and later gained an API.
  • Kompozy runs Avatar-IV-class avatar generation inside Persona Shorts and Persona HeyGen, then captions, reframes, repurposes, and publishes across platforms.

Frequently asked questions

What is HeyGen Avatar IV?

Avatar IV is HeyGen's image-to-video avatar model. It turns a single photo and an audio track (or script) into a talking, moving video with synced lips, expressive facial movement, and authentic hand gestures. It launched in 2025 and can work from tilted, profile, or angled source photos, in hyper-realistic or stylized styles.

What did HeyGen and Google Cloud announce on August 13, 2026?

A joint engineering write-up on the Google Developers Blog detailing how HeyGen ported the 18-billion-parameter Avatar IV model onto Google Cloud Trillium (v6e) TPUs. Using torchax, FSDP weight sharding, and Ulysses sequence parallelism, HeyGen reached a 1.86x speedup over its first working version and reports up to 25% lower cost per minute of generated video versus an 8×H100 setup.

Does running Avatar IV on TPUs make HeyGen cheaper for me?

Not directly or immediately. The 1.86x speedup and up to 25% cost reduction are HeyGen's serving costs, not the sticker price. Infrastructure savings like this typically show up later as more generous credit allowances or lower per-minute costs, but HeyGen's public plans and credit rates did not change on the day of the announcement.

How does a content engine like Kompozy fit avatar models like Avatar IV?

Avatar IV generates the talking-head video; it does not caption, reframe, repurpose, or publish it. Kompozy runs Avatar-IV-class avatar generation natively inside its Persona Shorts and Persona HeyGen formats, then auto-captions and reframes the clip, fans the idea into carousels, threads, blogs, and newsletters via a Persona Brief, and schedules it across the eight social platforms plus blog and email.

Related news

← All AI news · Get started →