SOCIAL INTELLIGENCE
Natural conversations. Consistent characters.
Built to understand context and emotion.
Il y a des idées, des sentiments qui ne surgissent ou ne se transforment en actes que chez les individus en foule.
There are ideas and feelings that arise or turn into action
only in individuals who are part of a crowd.
SOCIAL INTELLIGENCE
Trained on personality, emotion, and social context for consistent, natural dialogue across turns.
Parameters
Current release
SOCIAL INTELLIGENCE ARCHITECTURE
Social state, memory of people and events, and a sparse Transformer are trained together, connecting language, actions, and state updates in one pipeline.
View full architecture
ARCHITECTURE SPECIFICATIONS
Sparse Transformer
d_model
Top-8 active · 1 shared expert per layer
| Component | Configuration |
|---|---|
| Language backbone | 72-layer sparse Transformer |
| Hidden dimension | 8,192 |
| Routed experts per layer | 256, Top-8 active |
| Shared experts | 1 per layer |
| Intermediate size per expert | 1,536 |
| Attention | Gated GQA, combining local and global layers |
| Social state readout | Every 4 layers in initial experiments |
| Active state slots | 64–128 in initial experiments, allocated per participant |
| Memory | Trainable retrieval and cross-attention, across sessions |
| Outputs | Language, actions, state updates, future distributions |
| Training | Jointly optimized from pretraining onward, followed by interaction training |
CHARACTER & COMPANIONSHIP
Stay in character. Keep the conversation going.
For role-play, companionship, education, and mental health support apps. We train and evaluate OmniSoul for character consistency, emotional expression, and coherent multi-turn conversations.
EMOTION AND CONTEXT
We study how responses adapt as emotions, information, and relationships change, from recognizing feelings to expressing them naturally across turns.
SEMANTIC & MEMORY CONTINUITY
Track people, events, commitments, and timelines within the supplied context. Update the response as facts change, while keeping it consistent with earlier turns.
RESEARCH & MEASUREMENT
Evaluate the conversation.
Understand the whole interaction.
We evaluate personality, emotion, meaning, and outcomes together to study consistency across long conversations.
DIGITAL TWINS & POPULATION SIMULATION
From individual perspectives to group behavior.
OmniSoul uses synthetic personas to explore how different groups respond to the same situation.
PERSONA → OMNISOUL
Explore how identity, preferences, and experience shape individual decisions.
Scale of external persona resources8.3 billion persona records · MatrAIx Persona 8B
Sparse routed expertsActivated on demand, combined with shared experts for inference
Character context and memoryOn-demand sampling and batching keep simulation costs under control
CONDITIONS → DISTRIBUTIONS
Explore how character state and context shape the distribution of responses to an intervention.
Adjust the price to explore how adoption changes across groups.
Price-sensitiveIllustrative change in adoption propensity
Estimated at 700 tokens per persona, plus 180 input and 120 output tokens per turn. Input: 7,040,000, of which 4,200,000 can be reused. Actual API pricing and caching rules apply.
Charts and forecasts are interactive illustrations based on synthetic personas. 8.3B is the scale of the external research persona set; the public core set contains about 1 million. MatrAIx research
AGENT SOCIETIES
Explore what happens when agents interact.
Agents negotiate, cooperate, and compete in a shared environment. Each has its own goals, information, and relationships, and adapts to what others say and do.
We study how agents understand each other’s intentions, coordinate goals, and navigate conflict.
Explore researchUNDERSTANDING WITH RESPONSIBILITY
We envision people and agents living, learning, and creating together. Through shared experiences, their ideas can grow into collective action.