Performance

How fast Essence 2 and Expression 2 render on each platform.

× real time is seconds of avatar video rendered per second: at 1× an avatar keeps up with a live conversation, and anything higher is headroom.

Device	Essence 2	Expression 2
iPhone 18 Pro	4.9× real time, 124 fps	11.9× real time, 239 fps
iPhone 15	2.1× real time, 54 fps	5.5× real time, 111 fps
Galaxy S25+	2.0× real time, 52 fps	2.4× real time, 48 fps
Mac (M4)	6.9× real time, 174 fps	8.8× real time, 176 fps
Linux PC (Intel i7)	2.0× real time, 50 fps	2.4× real time, 48 fps
Windows PC (Intel i7)	0.6× real time, 17 fps	1.0× real time, 21 fps
Browser (Chrome, M5)	4.0× real time, 101 fps	2.4× real time, 49 fps
Cloud GPU (RTX 4090)	4.1× real time, 104 fps	17.0× real time, 340 fps
Cloud Mac (M4 Max)	2.8× real time, 71 fps	5.5× real time, 111 fps
Cloud CPU	1.1× real time, 28 fps	1.3× real time, 27 fps

Each figure is one avatar session rendering as fast as the device allows, on the fastest SDK we measured on that device. How we measure has the method, memory and the raw data.

All measurements

Every SDK and release measured, fastest first on each device.

Device	Model	SDK	Speed	fps	Live fps	First speech	Measured	Method
iPhone 18 Pro	Essence 2	Swift package 2.21.4	4.9×	124	25	0.38 s	2026-10-08	Essence 2 example app benchmark: avatar downloaded by agent code, unpaced x3 back to back, then a live 25 fps run; Release, published 2.21.4
iPhone 18 Pro	Expression 2	Swift package 2.21.4	11.9×	239	19.9	—	2026-10-08	examples swift/ios-expression2 (Release, 2.21.4), generate then play
iPhone 15	Essence 2	Swift package 2.17.3	2.1×	54	—	—	2026-09-27	slowest of 3 runs (model record)
iPhone 15	Expression 2	Swift package 2.18.0	5.5×	111	—	—	2026-09-27	slowest of 3 runs (model record)
Galaxy S25+	Essence 2	essence2-android 0.7.0	2.0×	52	—	—	2026-09-25	slowest of 3 runs (model record)
Galaxy S25+	Expression 2	expression2-android 0.4.10	2.4×	48	—	—	2026-09-23	slowest of 3 runs (model record)
Mac (M4)	Essence 2	Python · bithuman 2.11.12	6.9×	174	—	—	2026-09-25	slowest of 3 runs (model record)
Mac (M4)	Essence 2	Swift package 2.21.4	4.3×	109	24.5	0.40 s	2026-10-08	Essence 2 probe on the published v2.21.4 binaries (unpaced pull loop to the reply’s last frame; live 25 fps loop)
Mac (M4)	Essence 2	CLI 2.8.1	4.2×	106	—	—	2026-09-27	slowest of 3 runs (model record)
Mac (M4)	Expression 2	Swift package 2.21.4	8.8×	176	—	—	2026-10-08	examples swift/macos-expression2 (2.21.4, timed to the last frame), swift run -c release
Mac (M4)	Expression 2	Python · bithuman 2.11.12	8.4×	169	—	—	2026-09-25	slowest of 3 runs (model record)
Mac (M4)	Expression 2	CLI 2.8.1	8.4×	168	—	—	2026-09-27	slowest of 3 runs (model record)
Linux PC (Intel i7)	Essence 2	CLI 2.8.1	2.0×	50	—	—	2026-09-27	slowest of 3 runs (model record)
Linux PC (Intel i7)	Essence 2	Python · bithuman 2.12.1	1.2×	30	24.4	0.89 s	2026-10-08	notes/perf/python/bench.py (bithuman.open + Avatar.render, first request -> last frame) and live.py (AsyncBithuman, 40 ms chunks at real-time pace); warm-up excluded; CPU only, no cap
Linux PC (Intel i7)	Expression 2	Python · bithuman 2.12.1	2.4×	48	20	—	2026-10-08	notes/perf/python/bench.py (bithuman.open + Avatar.render, first request -> last frame) and live.py (AsyncBithuman, 40 ms chunks at real-time pace); warm-up excluded; CPU only, no cap
Linux PC (Intel i7)	Expression 2	CLI 2.8.1	2.2×	44	—	—	2026-09-27	slowest of 3 runs (model record)
Windows PC (Intel i7)	Essence 2	Python · bithuman 2.12.9	0.6×	17	—	—	2026-11-06	Python SDK on Windows, full-body avatar, 8 vCPU; the slowest of four runs is shown
Windows PC (Intel i7)	Expression 2	Python · bithuman 2.11.18	1.0×	21	—	—	2026-09-29	slowest of 3 runs (model record)
Browser (Chrome, M5)	Essence 2	web viewer dcba335	4.0×	101	25	—	2026-10-08	median of 3 runs
Browser (Chrome, M5)	Expression 2	web viewer dcba335	2.4×	49	—	—	2026-10-08	median of 3 runs
Cloud GPU (RTX 4090)	Essence 2	cloud API	4.1×	104	—	—	2026-09-27	slowest of 3 runs (model record)
Cloud GPU (RTX 4090)	Expression 2	cloud API	17.0×	340	—	—	2026-09-23	slowest of 3 runs (model record)
Cloud Mac (M4 Max)	Essence 2	cloud API	2.8×	71	—	—	2026-09-26	slowest of 3 runs (model record)
Cloud Mac (M4 Max)	Expression 2	cloud API	5.5×	111	—	—	2026-09-23	slowest of 3 runs (model record)
Cloud CPU	Essence 2	cloud API	1.1×	28	—	—	2026-09-25	slowest of 3 runs (model record)
Cloud CPU	Expression 2	cloud API	1.3×	27	—	—	2026-09-24	slowest of 3 runs (model record)

Only these devices are measured. Other hardware renders at other rates.
Browser is the avatar rendering in the visitor’s own tab with WebGPU (render=local). By default the avatar renders in the cloud and streams to the page.
Cloud is how fast one finished video is delivered, encoding included, on a server with no other sessions. bitHuman picks the server for each session.
A live conversation plays at 25 fps for Essence 2 and 20 fps for Expression 2. Extra speed makes a video file finish sooner; it does not put more frames on screen.

Setup for each SDK: Apple, Android, Web, Python, CLI.