Changelog archive

Essence 2 — sharper mouth and teeth, and a much smaller model file (2026-07-27)

Docs

• Essence 2 renders through a new unified renderer: a sharper mouth interior and teeth, and wider, re-centred mouth motion.
• The downloadable model is much smaller — read Content-Length rather than hard-coding a size. No API, tier or pricing change.
• New creations get it automatically; existing agents move over when retrained.

Expression 2 — faster creation, same quality bar (2026-07-22)

Docs

• Expression 2 creation runs an adaptive-ladder training recipe by default: essentially the same quality at a fraction of the training time and cost. No API, pricing or serving change.

Run Expression 2 locally from the CLI (2026-07-16)

CLI

• The bitHuman CLI renders expression-2 avatars on your own hardware (Apple Silicon macOS, Linux x86_64), with the render engine inside the CLI. bithuman run with no arguments renders the free Wise Pup avatar.

Expression 2 — smaller, sharper serving model (2026-07-16)

Docs

• Expression 2 serves each identity through a more compact model with a sharper mouth and teeth, on gallery identities and new creations. No API, contract or pricing change; no action needed.

Expression 2 — adaptive per-identity training (2026-07-15)

Docs

• Expression 2 creation adapts training per identity: every agent passes the same quality checks, and an identity that needs more work gets more training. No action needed. See Expression 2.

Agent creation is image-only (2026-07-10)

Docs

• Breaking: the video creation input is removed for all models. Provide a portrait image (or a prompt); a request carrying video gets 400 VIDEO_INPUT_NOT_SUPPORTED before anything is billed.
• video_aspect_ratio is removed and duration is ignored. Existing agents are unaffected, and file upload still accepts video as an asset.

Essence 2 naming settled (2026-07-10)

Docs

• essence-2 is the standard tier name — the former essence-2-light was consolidated into it on 2026-07-05 — and the premium tier is no longer offered publicly. See Renamed and retired names.
• Rates are unchanged. The model guide is at /models/essence-2; the old /concepts/essence-2-light and /concepts/essence-2-quality URLs redirect.

Expression 2 creation price: 2000 credits (2026-07-10)

Docs

• expression-2 creation and model-add cost 2000 credits, up from 500. Essence 2 stays at 500; v1 stays at 250.
• auto bills the routed model’s rate; GET /v1/pricing advertises auto at the 2000 ceiling.

Essence 2 & Expression 2 — launch rollout begins; model pages refreshed (2026-07-10)

Docs

• Creation access opens progressively: a v2 creation ahead of your account’s access returns 503 MODEL_NOT_YET_AVAILABLE and bills nothing.
• The essence-2 and expression-2 model pages gained their shipping characteristics — resolution, serving tiers and browser-local status.

Plan concurrency, offline licensing preview, and one pricing page (2026-07-10)

Docs

• Concurrent avatar sessions are a plan allowance; a session start beyond it returns 403 CONCURRENCY_LIMIT_REACHED, and live sessions are never cut off. See Session concurrency.
• Offline licensing was previewed; its terms have since changed — see Pricing → Offline licensing.
• Pricing is now the single home of every rate, and naming history lives under Models → Naming & migration.

Multi-agent avatar rooms — audio binds to the launching agent (2026-07-09)

Docs

• In a room with several agents, the cloud avatar binds its audio to the agent that started the AvatarSession (via lk.publish_on_behalf), not the first agent it sees. Server-side; no upgrade needed. See LiveKit → Multiple agents.

essence-2-light consolidated into essence-2; force-tier slugs (2026-07-05)

Docs

• The essence-2-light name is retired: create and render with model: "essence-2". The old name gets a 400 pointing at essence-2; existing agents and saved links keep working.
• New force-tier slugs (essence-2-gpu / -ane / -cpu, expression-2-gpu / -cpu / -ane) pin one serving tier and never overflow. See tier pinning.

Android / Kotlin SDK docs restored (2026-07-04)

Android

• The Android SDK page and the Kotlin hello-avatar example are back. ai.bithuman:sdk:2.3.6 (essence-1, arm64-v8a) was unchanged; only its documentation had been removed.

Pick-for-me creation, model adds & downloads (2026-07-02)

Docs

Named as of today: the tiers then called Essence 2 Light and Essence 2 Quality are now essence-2 and an Enterprise-only tier — see Renamed and retired names.

• model: "auto" on POST /v1/agent/generate routes a photorealistic person to essence-2 and a cartoon, animal or creature to expression-2. An explicit Essence 2 creation of anything but a photorealistic human gets 422 MODEL_SUBJECT_MISMATCH before billing.
• New: POST /v1/agent/{code}/models adds a model to an existing agent, and GET /v1/agent/{code}/model/download downloads one — bithuman pull <AGENT_CODE> uses it.

Official model guides + natural idle for the second generation (2026-07-02)

Docs

• Each second-generation model has its own guide — Expression 2 and Essence 2 — plus a session troubleshooting guide.
• Expression 2 idles on a looping clip of the identity itself, and idle loops always play forward.
• Agent responses carry supported_models; asking for a model the agent has not generated yet returns 409 MODEL_NOT_GENERATED before any charge.

Announced — Essence 2 & Expression 2 (launching July 10, 2026)

Docs

• essence-2 and expression-2 are announced for July 10, 2026 on every surface; essence-1 and expression-1 stay fully supported. See Essence 2 & Expression 2.

Talking video generation — new API (2026-06-29)

API

• New POST /v1/video/generate and GET /v1/video/{job_id} render a talking-video mp4 of your agent from text or audio, asynchronously. Launch engines are expression-2 and an Enterprise-only tier, billed per minute of output rounded up; a failed render is refunded.

Agent generation — v2 model names accepted (2026-06-29)

Docs

POST /v1/agent/generate accepts expression-2, essence-2-light and an Enterprise-only tier as model values, ahead of the v2 launch.

Update 2026-06-30: the pre-release aliases (elevate, embody, embody-gpu, essence-2-mobile) were retired and return a 400 VALIDATION_ERROR.

Model naming — versioned public taxonomy (2026-06-26)

Docs

Model families get versioned public names — essence-1, essence-2-light, expression-1, expression-2 and an Enterprise-only tier — on agent generation and the viewer. essence and expression map to essence-1 / expression-1.

The pre-release codenames were transitional aliases and have since been retired (see the 2026-06-29 entry). Share links are unaffected.

Python SDK bithuman 2.3.10 (2026-06-23) — self-hosted streaming lag fix

Python

• Self-hosted streaming holds a steady frame rate across a long turn, and the audio stream resets at each turn so idle frames cannot shift lip-sync.

Python SDK bithuman 2.3.9 (2026-06-23) — barge-in / interrupt fix

Python

• Interrupting the avatar no longer freezes the runtime after the first barge-in; a user can talk over it repeatedly.
• Recommended LiveKit stack: bithuman 2.3.9+ with livekit-plugins-bithuman 1.6.3 and livekit-agents 1.6.x.

Python SDK bithuman 2.3.8 (2026-06-16)

Python

• Maintenance release on the 2.3 line (2.3.5–2.3.7 were not published).

Python SDK bithuman 2.3.4 (2026-06-12) — Linux CA auto-discovery

Python

• On Linux the SDK finds your distro’s CA bundle itself, so self-hosted auth needs no symlink workaround. CURL_CA_BUNDLE / SSL_CERT_FILE still win when set — unset them unless they point at a valid bundle.
• The macOS wheels require macOS 26+ (arm64).

2.3.0 (2026-05-28) — layered architecture + PyPI wheel split

Docs

• pip install bithuman is now the Python library only; install the CLI with Homebrew or the universal installer (the CLI page). The Python and Swift APIs are unchanged.
• The CLI keeps run, render, info, pull, list, doctor and init; legacy 1.x verbs stay removed.
• BITHUMAN_BRAIN_* env vars are renamed BITHUMAN_AGENT_* (the old names still work, with a warning). bithuman.utils and bithuman.audio are removed. Engine ABI v7 adds be_runtime_tick_compose_from_mel; older builds keep working.

Python SDK bithuman 2.2.2 (2026-05-25) — Linux CLI tarballs restored

CLI

• A CI-only release: the Linux CLI tarballs are back on the GitHub Release. No API or runtime change.

Python SDK bithuman 2.2.1 (2026-05-25) — the on-device brain

Python

• 2.2.0 was skipped; install 2.2.1.
• BITHUMAN_LOCAL=1 gave bithuman run a fully on-device conversation brain (whisper.cpp, llama.cpp, Supertonic TTS, Silero VAD) — no OpenAI key, no network. The wheel that carried it is gone; see local mode.
• New plugins: livekit.plugins.bithuman.{WhisperSTT, LlamaCppLLM, SupertonicTTS}.

Python SDK bithuman 2.1.0 (2026-05-24) — figure → avatar

Python

• --figures-root is renamed --avatars-root (the old flag stays as a deprecated alias), and the default cache moves to ~/.cache/bithuman/avatars. No runtime change.

Python SDK bithuman 2.0.2 (2026-05-24) — graceful drain

Python

• bithuman run waits for the engine to tear down before exiting, fixing a crash (exit 134) on Ctrl-C or a supervisor stop.

Python SDK bithuman 2.0.1 (2026-05-24)

Python

• AsyncBithuman.cleanup is async — await b.cleanup() works.
• CLI error messages and essence-render --help no longer name renamed commands.

Python SDK bithuman 2.0.0 (2026-05-22) — bundled-CLI release

CLI

• pip install bithuman ships a bithuman command that runs the full talk-to-your-avatar stack. The library API is unchanged; the 1.x Python CLI lives on as essence-render.
• Wheels: macOS arm64, Linux x86_64 and aarch64; Python 3.10+.

v1.18.5 (2026-05-18)

Docs

• One pip install bithuman carries the full 1.11.3 API plus the native engine, backward-compatible. Python 3.9–3.14; pin >=1.18.5 (Windows and Intel Mac stay on ==1.11.3).

v1.17.x (2026-05-14)

Docs

• bithuman avatar --openai renders an avatar in the browser on OpenAI Realtime; voice / text auto-pick cloud or --local.
• An interactive TUI for voice (BITHUMAN_NO_TUI=1 opts out).
• The Flutter plugin is renamed bithuman_avatar → bithuman.

v1.16.0 (2026-05-14)

Docs

• Streaming API on Swift (pushAudio/frames()/resetStream()).
• Default Realtime model: gpt-realtime-mini.

v1.12.0 (2026-05-12)

Docs

• First unified release: Python, Swift and CLI from one source, identical output.
• Linux and Windows Python wheels (no WSL).