Actra — Action
What should happen?
LabsAI · Research Brief · Model Families
The Labs model families for speech that acts, orchestration that listens, and visualization that manifests.
September 2026
An Actra converts speech and world state into action. An Octra converts objective and resources into orchestration policy. A Vistra converts speech, intent, context and state into a visual representation. They are the model families behind Speech‑to‑Action (STA), Speech‑to‑Orchestration (STO), and Speech‑to‑Visualization (STV) — three competences a spoken intelligent system needs the moment it stops merely talking and begins acting, coordinating, and manifesting information into interactive visual form.
Action-Centered Transducers for Reasoning and Agentic Systems · singular: an Actra
Implements the STA class — Speech-to-Action
An Actra takes speech, conversation, application state, memory, intent and the vocabulary of available actions and produces action: what the system does, co-timed with what the intelligence says.
The name is a transducer’s name. It describes the conversion performed rather than prescribing the topology that performs it, because the family is architecture-agnostic by design.
Actras are to speech what VLA models are to vision-and-action — and in the terms that made the previous generation of models legible, an Actra is a GAT: Generative Action Transducer.
Its generative object is not principally the sentence. It is the action.
Orchestration-Centered Transducers for Routing and Agentic Systems · singular: an Octra
Implements the STO class — Speech-to-Orchestration
An Octra takes state, objective and resources and produces orchestration policy: model and tool selection, decomposition, parallel versus sequential execution, escalation, verification paths, recovery and termination.
In the spoken loop, all of it can be steered live by speech.
STA decides what the system does. STO decides which intelligence does it and how the intelligences work together.
An Octra is therefore a GOT: Generative Orchestration Transducer — generating the coordination itself rather than merely generating words describing coordination.
What it generates can run as an AI Orch: the executable coordination object binding models, agents, tools, workflows, memory, policies, data and compute into one live operation.
Orch Models are intelligence. AI Orchs are execution structures.
Visualization-Centered Transducers for Representation and Agentic Systems · singular: a Vistra
Implements the STV class — Speech-to-Visualization
A Vistra takes speech, conversation, intent, world state, application state, available information, visual context and representational constraints and produces visualization.
Not merely an image. Not merely a chart. Not merely generated UI.
A visualization can be a chart, interface, diagram, simulation, spatial composition, live webpage, media surface, data environment, interactive object, generated scene, evolving canvas or a composition of several of these at once.
The Vistra’s prediction problem is: What should this become visually?
STV therefore treats visualization as a native generative target rather than a downstream formatting operation performed after language generation.
A Vistra is a GVT: Generative Visualization Transducer.
Its output can describe the visual object, its structure, its spatial relationships, its state, its affordances, its timing, its relationship to speech and the transformations that should occur as the interaction continues.
Where conventional multimodal systems may generate language about what they see, STV establishes the inverse capability: speech becomes visualization.
The spoken request is not merely answered. It is manifested.
The three families form a designed system.
What should happen?
What intelligences, tools and resources should make it happen, and how should they coordinate?
What should become visible, and in what representational form?
These are related but irreducible prediction problems.
A user can say: “Show me how NVIDIA’s revenue changed over the last decade, explain the inflection points, compare them with the AI infrastructure cycle, and let me manipulate the chart while we’re talking.”
An Octra can determine which models, retrieval systems, data sources and tools should participate. An Actra can retrieve the information, operate the interface, change states and respond to subsequent commands. A Vistra can determine that the information should become an interactive temporal visualization, construct its representational structure, synchronize its appearance with the explanation and transform it as the conversation changes.
The system does not have to reduce the entire operation to a stream of words. It can coordinate, act and visualize.
Speech-to-Visualization is the competence of transforming spoken intent directly into an appropriate visual representation while preserving the semantic, temporal and interactive relationships of the conversation. We abbreviate it: STV.
STV is not simply text-to-image with speech transcription placed in front of it. And it is not simply voice-controlled generative UI. Those systems can be components of an STV system, but neither defines the class.
The distinction is the prediction target. For an STV system, visualization itself is the target. The system must infer not only what the speaker said but what visual representation best carries what the speaker means. That requires decisions about representation.
A comparison may become a chart. A hierarchy may become a graph. A physical concept may become a simulation. A process may become an animated diagram. A collection may become an explorable spatial interface. A website may become a live shared surface. A concept that does not yet have a predetermined interface may require the interface to be generated. And a conversation may require several of these representations to coexist and transform over time.
The visualization is therefore not decoration attached to an answer.
It is part of the answer’s intelligence.
The first-generation LabsAI Studio already demonstrates an important precursor to this model class: charts, tables, cards, media and live web surfaces can be composed into a shared visual stage during the sentence explaining them. The introduction of the Vistra family makes the research direction explicit.
The long-term objective is not merely for a language model to request predetermined UI components. It is for visualization decisions themselves to become learned model behavior.
The shared stage becomes a stateful visual environment. Objects on that stage can appear, disappear, transform, expose affordances, respond to speech, participate in actions and become inputs to subsequent reasoning.
This changes the conversational loop. The system does not simply remember what was said. It can maintain knowledge of what currently exists visually.
Speech can refer to that state: “Move that over there.” · “Expand the second one.” · “Turn this into a timeline.” · “Show me what happens if this doubles.” · “Keep the chart, but replace the table with the underlying companies.” · “Visualize what you’re explaining.”
Those utterances are not requests for another paragraph. They are instructions against a living visual state.
The resulting spoken-intelligence stack contains three native transformations:
Speech‑to‑Action
Speech → Action
Implemented by Actras.
Speech‑to‑Orchestration
Speech → Orchestration
Implemented by Octras.
Speech‑to‑Visualization
Speech → Visualization
Implemented by Vistras.
The shortest form of the architecture becomes:
Actras act. Octras coordinate. Vistras manifest.
Or, expressed as prediction targets:
Actra → Action
Octra → Orchestration
Vistra → Visualization
Language can participate in every transformation without defining any of them.
Charts, tables, cards, media and live web surfaces can be composed into a shared visual stage while the sentence explaining them is still being spoken. This is the first-generation bridge toward STV: visualization occurring inside the conversational loop rather than after it.
Elements can expose voice-addressable affordances for as long as they exist. “Pause it.” “Fold the second one.” “No — the other one.” The system resolves those expressions against visual state rather than treating them as isolated text.
STV extends that interaction from manipulating existing objects toward generating and transforming representation itself. “Make that a graph.” “Show this geographically.” “Turn those numbers into something I can understand visually.” “Build me an interface for this.” “Visualize the relationship.” The target is no longer merely the correct action against an existing interface. It is the correct visual form.
An action is verified, attempted, blocked or failed against the actual interface state. The same epistemic requirement applies to visualization. A system should distinguish between intending to create something and successfully placing it into the shared visual environment.
A real page can become part of the shared stage, allowing the conversation to ground itself in what is actually rendered.
Retrieval, generation, tools and orchestration can execute while the spoken interaction remains alive. This remains the Octra layer’s floor discipline.
The addition of STV means the system can simultaneously maintain three evolving states:
conversation state, execution state and visual state.
Every interactive session can leave more than an action trajectory. It can leave a visual trajectory.
What was requested. What information was available. What representation was selected. What appeared. Where it appeared. When it appeared relative to speech. What the user referenced. What the user changed. What representation failed. What representation was replaced. What the user preserved. What the system learned from the correction.
Those trajectories form a candidate native corpus for Vistras.
The objective is not to train a model merely to reproduce attractive interfaces. The objective is to learn the relationship:
intent → representation → interaction → outcome.
A successful visualization is therefore not necessarily the prettiest visualization. It is the representation that most effectively carries the user’s intended understanding and remains useful as the interaction evolves.
As with Actras and Octras, candidate improvements should be evaluated offline and promoted deliberately. Nothing about defining a learned visualization class requires uncontrolled self-modification in production.
Within LASIMA — Labs Artificial Super Intelligence Model Archetypes, LaLaMo, LaMimo and Lasamo remain hierarchical forms rather than separate model families. Each of the three families can therefore exist at each LASIMA scale.
| Family | LaLaMoLabs Large Models | LaMimoLabs Medium Models | LasamoLabs Small Models |
|---|---|---|---|
| ActrasSpeech‑to‑Action | Frontier-scale action reasoning and execution | Surface-tuned working-scale action | On-device and low-latency action |
| OctrasSpeech‑to‑Orchestration | Full-fleet orchestration across Labs intelligences | Session-scale orchestration | Embedded routing and floor discipline |
| VistrasSpeech‑to‑Visualization | Frontier-scale multimodal visualization and dynamic world representation | Surface and application-scale visualization | Local, low-latency and device-native visualization |
The addition of Vistras therefore does not create another size class inside LASIMA. It adds another prediction family across the LASIMA ladder.
The distinction matters. LaLaMo, LaMimo and Lasamo answer: At what model scale does this intelligence operate? Actra, Octra and Vistra answer: What does this intelligence natively predict?
The broader research argument is not that language becomes unimportant. It is that language is too important to confuse with intelligence itself.
Speech can initiate reasoning without requiring language to remain the final computational object. A spoken instruction can become an action. A spoken objective can become an orchestration. A spoken idea can become a visualization. And increasingly, those outputs can feed one another.
Speech → Orchestration → Action → Visualization
Visualization → Speech → Action
State → Orchestration → Visualization → Action
The loop is multimodal, stateful and recursive. That is the larger direction behind the three families.
We are interested in models whose native outputs correspond increasingly closely to the things intelligent systems actually need to produce in the world. Not only words.
Actions. Orchestrations. Visualizations.
The first-generation Actra surface and Octra layer are live in LabsAI Studio today. The visual behaviors already present in that system provide the production substrate and interaction record from which the STV research direction emerges.
Vistras and Speech‑to‑Visualization extend that work into a formal third model family.
The research program now asks a larger question: If language models learned to generate language, and Actras learn to generate action, and Octras learn to generate orchestration —
what happens when models learn to generate the visual manifestation of thought itself?
That is the problem Vistras are intended to explore.
FeMa, the Frontier Editorial Media Architecture, is the first program planned to draw on these families outside the Studio, for editorial and media work in FrontierAI and Frontier Journal. It is not a further family and adds no rung to the grid above; the LaLaMo, LaMimo and Lasamo forms stay with the families themselves. FeMa‑1, its first implementation, is in development.
The argument these families exist to prove is stated in the companion note.
Read Larger Than Language →Contributing authors: Maya E. Davis · Duránd F. Davis Jr.
This research is published while the questions are still open, and the systems it describes are live. We would love for you to join us — and please share your thoughts at research@labsintelligence.ai.
Please cite this work as:
Davis, Maya E., and Davis, Duránd F., Jr., “Introducing Actras, Octras & Vistras.” LabsAI Research Briefs, Labsintelligence — lab1 of Labs Companies, Inc., September 2026.
Or use the BibTeX citation:
@article{labsintelligence2026introducingactrasoctras,
author = {Davis, Maya E. and Davis, Duránd F., Jr.},
title = {Introducing Actras, Octras \& Vistras},
journal = {LabsAI Research Briefs},
publisher = {Labsintelligence, lab1 of Labs Companies, Inc.},
note = {LabsAI and its sister FrontierAI are built by Labsintelligence on LABA, LabsAI for LabsAI Studio and FrontierAI for Frontier Journal},
year = {2026},
month = {september},
url = {https://research.labsintelligence.ai/labsai/introducing-actras-octras/},
}