Byte 3.0 Realistic Control
Photorealistic procedural face + thinking machine. First load caches model + 3.6MB brain.
Local model — routing by task
SmolLM2 360M · fastest · tool specialist
SmolLM2 1.7B · ToT+Reflexion balanced
Llama 3.2 1B · Reflexion + Constitutional
Llama 3.2 3B · full research stack
Qwen2.5 0.5B · multilingual fast CoT
1.7B sweet spot for ToT + Reflexion. Face is pure canvas — no model needed for avatar.
Load thinking model
Unload
No model loaded — face + retrieval live
Realistic face renders without model. Model adds LLM-powered ToT, Reflexion, Constitutional AI reasoning.
Reasoning + tools
1126 · ToT depth 3
Reasoning mode — research stack
Auto — routes to best (ToT/Step-Back/CoT)
ToT — Tree-of-Thought BFS (Yao 2023)
Self-Consistency vote 3 paths (Wang 2022)
Reflexion self-correct (Shinn 2023)
Step-Back principle (Zheng 2023)
Least-to-Most decompose (Zhou 2022)
CoT Chain-of-Thought (Wei 2022)
ReAct Reason+Act (Yao 2022)
Face controls — FACS action units
Skin: light Skin: medium Skin: dark Skin: cool undertone
Eyes: blue Eyes: hazel Eyes: brown Eyes: green
Emotion: neutral (AU0) Happy: AU6+AU12 Curious: AU1+AU2+AU5 Thinking: AU4+AU7 Concerned: AU1+AU4+AU15 Excited: AU1+AU2+AU12+AU25
Hair: none (focus face) Hair: short procedural Hair: long
TTS lip-sync
Show ToT + traces
Private memory — GenAgents importance+recency
remember (imp scored) forget
Realistic face tech: Procedural Canvas 2D, no images. Skin: radial gradient + subsurface scattering + noise texture. Eyes: sclera gradient, iris radial with texture lines, pupil, 2 highlights, eyelids with blink (0.14s), micro-saccades (0.5-2s random), eye tracking via mouse (FACS AU). Eyebrows: bezier with hair strokes, FACS AU1 inner brow raiser, AU2 outer brow raiser, AU4 brow lowerer. Mouth: lips gradient, teeth, tongue, FACS AU12 lip corner puller, AU25 lips part, AU26 jaw drop — lip-sync via TTS amplitude → viseme → mouth open. Emotions via FACS combinations. All 60fps.