AZI
Built to talk, remember, react, see, and become a little more alive.
SHE THINKS LOCALLY.
Azi runs her own local language model, giving her control over personality, timing, behavior and experimentation without turning every interaction into a remote API request.
SOMETIMES THE SYSTEM HAS TO SHUT HER UP.
Filtering happens dynamically between generation and output instead of flattening Azi’s entire personality. Hover a word to see what happened to it.
INTERCEPTOR also reads the first 160 characters of every reply. Some topics end the sentence before it is ever spoken — she gets a canned line instead.
NOT AN EMOTE SWITCH.
Emotions are states the model can combine, not a single reaction selected from a dropdown. Drag the faders.
25 FACES × 3 INTENSITY TIERS
20 DELIVERIES · 18 SOUNDS
CONTEXT BECOMES BEHAVIOR.
Before a message reaches her model, the system patches in what it knows: mood, the person, the room. Hover a tag.
SHE DOESN’T TREAT EVERYONE THE SAME.
Azi can build persistent relationships with individual users and use those relationships as context in future conversations.
your developer. you give him a hard time and trust him anyway.
TEXT ISN’T THE PERFORMANCE.
Fish Audio Drama 3 Preview turns Azi’s generated response into an expressive voice with far more room for character.
THE CONVERSATION DOESN’T RESET.
Persistent, structured memory lets Azi bring useful pieces of the past into future interactions. Hover a tape.
THE STREAM IS A DISTRIBUTED SYSTEM.
Overlays live as an independent module and communicate with Azi through a bridge that also carries Twitch and Discord context.
AND YES — SHE CAN SEE.
A VLM gives Azi visual context from the world around her, allowing what she sees to become part of the conversation.
- CONTINUE
- NEW GAME
- OPTIONS
- QUIT
FIVE VERSIONS LATER.
LONER puts every Azi through the same seven tests. This is how far she has come since V1.
| TEST | V1 | V2 | V4 | V5 | vs V4 |
|---|---|---|---|---|---|
| Talking | 13.4% | 42.1% | 53.6% | 68.3% | +14.7 |
| Following instructions | 2.6% | 33.4% | 49.1% | 65.9% | +16.8 |
| Voice | 32.7% | 36.1% | 63.9% | 82.3% | +18.4 |
| Vision | 0% | 0% | 30.6% | 51.2% | +20.6 |
| VTT | 9% | 21.4% | 62.5% | 71.2% | +8.7 |
| Interactions (Emotions) | 0% | 12.3% | 42.1% | 75.4% | +33.3 |
| Logic Path Building | 0% | 12% | 29.1% | 43.7% | +14.6 |
AZI AI / V5
Still experimental.
Still changing.
A little more alive every version.