Today we are launching the fifth generation of the engine behind MeshLite, which we call the Todalabs Engine. It is live now: the avatar on meshlite.io runs on it.
A lifelike avatar drawn in a data centre costs money for every minute it talks, and every extra conversation needs more graphics-card time. MeshLite draws the face on the visitor’s own phone or laptop instead. That changes how long a conversation can run, what a busy hour takes, and what it takes to start.
Long conversations, no per-minute fee
The face is drawn on the visitor’s device. A ten-minute conversation with a server-rendered avatar streams roughly 128 MB of 720p video; a MeshLite avatar downloads about 6 MB once, and after that only the voice travels.
Because no server of ours draws the face, MeshLite charges no per-minute fee for it. Within the capacity agreed in the licence, a conversation runs as long as your user needs.
Built to scale
Every conversation brings its own device to draw the face. Adding conversations adds no graphics-card load on our side for the face, so a launch-day spike doesn’t queue behind a GPU. Plans are sized by how many sessions run at once; growing means the next plan or a pack, never a per-minute charge. Voice and the language model run on providers you choose and scale on their terms.
Fast, and in sync
Talking in under 3 seconds. From opening the page to talking now takes under 3 seconds, down from under 5 seconds on our previous engine.
A mouth locked to the voice you hear. The mouth stays within 25 milliseconds of the sound the listener actually hears, measured on a desktop with and without a simulated Bluetooth delay. The human eye does not notice an offset below about 45 milliseconds (ITU-R BT.1359).
A voice AI that fits in under a megabyte. Our own voice AI reads the speech on the visitor’s device and picks the mouth shape for each moment. It has 1.76 million parameters per language, about 180 times fewer than the reference model it learned from, and it analyses ten seconds of speech in 68 milliseconds on an ordinary laptop, without a graphics card.
One photo to start
Until now, a lifelike avatar started with about five minutes of video and an hour and a half of processing. Now it starts with one photo. Upload a portrait and the engine prepares the avatar in about twenty minutes, in four attitudes: idle, listening, thinking and speaking. It moves between them in about a third of a second.
That is the one time a graphics card works: for about twenty minutes, when the avatar is created. Everything comes out of our own engine: the mouths, the videos and the voice AI that drives the lips. No external AI video service is involved at any step.
Before and after
| Before | Now | |
|---|---|---|
| What you provide | About 5 minutes of video | One photo |
| Time to create an avatar | About 1 h 30 | About 20 minutes |
| Voice AI download, per language | 8.6 MB | Under 1 MB |
| Time from opening the page to talking | Under 5 s | Under 3 s |
| Graphics cards drawing the face during the conversation | 0 | 0 |
How we got here
The Todalabs Engine is our fifth engine in four months. Each generation moved one more piece of the work off the data centre and onto the device.
- V1 · Fluide (July 2026). The face stopped being rendered during the call: every mouth is prepared once, then played back on the device in time with the voice.
- V2 · MeshLite (July–August). One line of code puts the avatar on any website, in French, English and Chinese, on desktop, Android and iPhone.
- V3 · Personalities (September). One avatar, several personalities, each with its own voice and language.
- V4 · Realism (October). Lifelike faces that breathe and look at you, and a start time cut from 11–14 seconds to under 5 in two days.
- V5 · Todalabs Engine (October). One photo, and everything from our own engine.
Why we went lighter
Research reports in this category describe avatar models getting bigger: one AI avatar provider describes a model of more than 18 billion parameters, running in a data centre for every minute someone talks to it. We went the other way. Shipping pixels hides how heavy a model is. Shipping the model does not, so ours has to be light.
Speech recognition, the language model and the voice run on servers from providers you choose, with their own costs.
Sources: dates from our development history, July–October 2026. Measurements from our October 2026 test benches; analysis speed measured on a laptop. Video comparison: a 720p stream at 1.7 Mbit/s, the public preset of a common video-call toolkit; an estimate at equal voice traffic. Perception threshold: ITU-R BT.1359. The 18-billion-parameter figure is that provider’s own published description. Session length and plan sizing: MeshLite public plans.