How we measure
How the frame rates are measured, on which releases, with the raw data.
Method
Measured in September 2026 on CLI 2.8.1, bithuman 2.11.12, bithuman 2.11.13, bithuman 2.11.18, Swift package 2.15.0, Swift package 2.17.3, Swift package 2.18.0, essence2-android 0.7.0, expression2-android 0.4.10, the hosted web viewer and the cloud API.
Frames per second from speech audio in to finished video frame out, for one avatar session, rendering as fast as the hardware allows on a reference speech clip. × real time is fps divided by the model’s playback rate (Essence 2 25 fps, Expression 2 20 fps); at 1.0× or more the avatar keeps up with a live conversation. Every figure is re-measured at least every 30 days on the releases listed above. Raw records, including the clip, per-run dates and memory: performance.json.
The macOS CLI Essence 2 figure was measured with BITHUMAN_THREADS=8; by default the CLI uses one thread per CPU it may use, up to 16. The Linux CLI uses default settings.
Memory
| Runs on | Hardware | Essence 2 | Expression 2 | Measured on |
|---|---|---|---|---|
| macOS · CLI | Apple M4 | 1.0 GB | 0.9 GB | CLI 2.7.8 |
| macOS · Python | Apple M4 | 1.2 GB | 0.8 GB | bithuman 2.11.12 |
| Linux · CLI | Intel Core i7-13700F (x86_64) | 2.2 GB | 1.6 GB | CLI 2.7.8 |
| Linux · Python | Intel Core i7-13700F (x86_64) | 1.0 GB | 1.5 GB | bithuman 2.11.12 |
| iPhone · Swift package | iPhone 15 | 0.9 GB | 0.5 GB | Swift engine 1.10.0, Swift package 2.14.1 |
| Android | Samsung Galaxy S25+ | 1.0 GB | 2.3 GB | essence2-android 0.7.0, expression2-android 0.5.0 |
| Web browser (WebGPU) | Chrome on Apple M4 | 3.0 GB (Chrome tab memory footprint) | 1.8 GB (Chrome tab memory footprint) | — |
Peak memory (RAM) of one render, or of one session on a phone, on the release in the last column. Cloud API: nothing to provision.
Time to a finished video
| Runs on | Hardware | Essence 2, first run | Essence 2, after that |
|---|---|---|---|
| macOS · CLI | Apple M4 | 4.0 s | 4.8 s |
Seconds from starting the command to a finished MP4 of a 10-second clip, for the whole process: loading, rendering, encoding and exit, timed on CLI 2.7.8. Each number is the slowest of three runs. “First run” is a fresh install with the model files already downloaded. “After that” is every later render on the same machine. Lip sync is checked on every run: no run published here plays more than 40 ms off the audio.
Later renders are not faster: the model files are already downloaded on the first run, so both columns measure the same work, within run-to-run variation.
Pin a tier for a benchmark
To measure one tier, append ?model= with a tier slug to the viewer or embed URL:
https://www.bithuman.ai/embed/A23WJF0199?model=expression-2-apple| Model | Tier slugs |
|---|---|
essence-2 | essence-2-gpu · essence-2-apple · essence-2-cpu |
expression-2 | expression-2-gpu · expression-2-apple · expression-2-cpu |
- A recognized slug pins the session to that tier: if the tier is unavailable, the session fails rather than playing elsewhere.
- An unrecognized slug is ignored and the session plays as usual. If a pin seems to have no effect, check the spelling.
- To be told about a typo, set the embed token’s
modelfield instead: an unknown value is refused with a400listing the accepted names when you mint the token.
In production, omit ?model= and let the service choose.
Raw data
performance.json holds every published cell: frame rate, × real time, release, date, the speech clip it was measured on, and memory. The pages under Performance are generated from it.