HLXION v3.1
Deep reasoning — science & code
Named for the double helix, HLXION is built for problems with a correct answer: proofs, protein-folding literature, production codebases. It shows its work and checks it twice.
iViRTH presents the DoubleSlit Project
Every answer starts as a wave of possibilities. YNTR is how it collapses into one.
A family of AI models built by iViRTH — and the physical products they'll eventually run inside, from public benches to a small robotic presence. Ask something, and instead of one scripted reply, YNTR resolves a whole spread of possible answers into the one that actually fits you.
The measurement problem, solved for language.
In the original double-slit experiment, a single particle of light passes through two slits at once, existing as a spread of possibilities until something measures it — at which point it resolves to one point on the screen. It's an odd way to describe physics. It turns out to be a useful way to describe a language model.
Every token a YNTR model produces starts the same way: as a wide distribution over what could come next, not a single scripted line. Your prompt, your history, your intent — that's the measurement. It's what decides which of the many possible responses actually reaches your screen.
A model that only ever gives one answer was never really thinking it through.
That's also where the harder work lives. We spend as much effort on how a model narrows its possibilities as we do on the possibilities themselves — training it to collapse toward answers that are honest, useful, and safe, even when nobody's checking.
Six models. One spectrum.
Every YNTR model sits somewhere between two instincts: precise, structured, particle-like reasoning on one end, and broad, exploratory, wave-like reasoning on the other. Pick the one that matches the shape of your problem.
VISION v2.2
Multimodal perception
VISION is built around seeing first — screenshots, charts, handwriting, live video — described and reasoned over as naturally as text, in the same conversation.
iON v3
Flagship general intelligence
The flagship of the family. Long-context reasoning, tool use, and conversation that holds up across a full working session, not just a single clever reply.
UNION v1.4
Unified multimodal
Text, image, audio and code in a single context window. UNION reads a whiteboard photo and a spec document in the same breath, and doesn't lose either thread.
SYMBION v1.2
Agentic & collaborative
Built to work alongside other models and tools rather than instead of them — a coordinator that plans a task, delegates the pieces, and checks its own output.
MINIONS family
Small, fast, everywhere
Not one model — a small, distilled family that runs on a phone, a browser tab, or a Raspberry Pi, for the questions that don't need a flagship to answer.
Beyond the model
Where the intelligence meets the pavement.
YNTR lives in an API. iViRTH also designs and builds the physical objects it will eventually run inside — starting with public furniture and a small robotic presence for shared spaces. Two products, on two very different timelines.
Shipping — v1.0
Smart Solar Bench
Solar-powered public seating
A public bench that generates its own power, lights itself after dark, and carries a digital display for wayfinding or advertising — built solid enough for a bus shelter, not a lab demo.
Solar canopy
Digital display
v2.0 adds AI
In development
Robotic Totem
An interactive presence for shared spaces
A small robotic totem designed to notice people, hold attention, and surface information — the character-driven counterpart to the bench's quiet infrastructure. Currently in concept and prototyping.
Concept stage
Prototyping
Sensing + interaction
Hardware · 01
Smart Solar Bench
Public seating that pays its own power bill. The Smart Solar Bench pairs solid timber-and-steel seating with a solar canopy, ambient lighting, and an integrated digital display — designed to drop into bus shelters, transit corridors, campuses, and plazas without new trenching or wiring.
Concept render — roadside installation with canopy and ad display
Solar canopy
A canopy roof does double duty — weather cover and power generation — so the bench runs without a mains connection.
Integrated display
A weatherproof digital panel built into the frame, for wayfinding, transit info, or ad space.
Ambient lighting
The canopy underside lights the seat after dark, keeping the stop visible and usable at night.
Modular installation
Bolts to a standard concrete pad. Deploys as a standalone bench or fitted into an existing shelter.
Built for public space
Steel frame and solid timber seating, engineered for continuous outdoor use — not a prototype finish.
Zero-dig power
A self-contained solar and battery system means no new electrical trenching to install one.
From bench to system.
v1.0
The bench itself
Built — shipping nowSolar canopy, ambient lighting, integrated display, and public-grade construction. This is the version being installed today.
v2.0
The AI layer
In developmentPresence sensing, adaptive display content, a usage dashboard for facilities teams, and predictive maintenance alerts — intelligence layered onto the same hardware.
v2.0's AI layer is under active development. Everything above the v1.0 build is a roadmap, not a shipped feature yet.
Hardware · 02
Robotic Totem
Where the Smart Solar Bench is quiet public infrastructure, the Robotic Totem is meant to be noticed. It's an upright robotic companion for shared spaces — part signage totem, part small robotic presence — designed to sense who's nearby, hold attention, and surface information at a glance.
Early concept — not final
Early concept render — full form study
Presence & interaction
Sensors that notice when someone's nearby and respond, instead of sitting inert until touched.
At-a-glance information
Wayfinding, alerts, or local information surfaced the moment someone looks up.
Expressive movement
Small articulated gestures in place of a static screen — a presence, not just a panel.
Part of the same network
Designed to sit alongside Smart Solar Bench installations in the same public spaces, not as a one-off object.
Development stage.
01
Concept & design
DoneForm, character, and the first component breakdown of the totem's core parts.
02
Prototyping
In progressBuilding and testing the physical components shown here — the current stage of work.
03
Field trials
UpcomingTesting an early unit in a real public space, alongside a Smart Solar Bench installation.
04
Public launch
UpcomingWider deployment, once field trials confirm the totem holds up outside the workshop.
The renders here are early concept and prototyping studies, not a final production design. Form and features will change before launch.
How a YNTR model gets built.
Three stages, run in order, every time — from a model that knows a little about everything to one you'd actually trust with a real question.
01
Superpose
Pretraining across broad, licensed and public data so the model holds many possible continuations at once, not a narrow script.
02
Interfere
Feedback-driven fine-tuning reinforces the response patterns people actually want and cancels out the ones they don't — like waves adding up or cancelling out.
03
Collapse
At inference time, your prompt is the measurement: the model resolves its possibilities into the single response you see, inside guardrails set well before you typed anything.
Understanding the model, not just the output.
Two threads of research run under every model release, and neither one is optional.
Interpretability
We build tools that let us look inside a running model while it works, rather than judging it only by what comes out the other end. If HLXION reasons its way to an answer, we want to see the path it took, not just the destination — and catch it quietly reasoning toward the wrong one.
Alignment & safety
Every YNTR model goes through a separate round of training whose only job is to narrow the collapse toward outcomes a person would actually want: decline the harmful request, show its work on the hard one, say "I don't know" when that's the honest answer.
Built for people who build.
One API for every model in the family. Switch from a MINIONS model to iON by changing a single string.
- Python, JavaScript and REST SDKs
- Streaming responses and batch inference
- Prompt caching for long-running context
- Usage-based pricing, no seat minimums
import yntr client = yntr.Client(api_key="sk-yntr-••••••••") # any model in the family, one call shape response = client.messages.create( model="ion-3", max_tokens=1024, messages=[ {"role": "user", "content": "Explain the double-slit experiment to a curious 12-year-old."} ], ) print(response.content)
Built for how the world actually speaks.
Most releases treat every language after English as an afterthought. YNTR models are trained on high-resource and low-resource languages together from the start — including the code-switching that shows up in real group chats, support tickets, and classrooms across India and beyond.
EnglishHindiGujaratiTamil
BengaliTeluguMarathiKannada
MalayalamPunjabiUrduOdia
ArabicMandarinJapaneseSpanish
FrenchPortugueseSwahili+40 more
The DoubleSlit Project.
iViRTH is the lab behind YNTR — and behind the hardware it will one day run inside. The DoubleSlit Project is our name for the long effort behind it: build models that hold real uncertainty instead of faking confidence, and give people an actual say in how that uncertainty resolves.
We called the model line YNTR — said the same way as "enter" — because every conversation is a doorway into that superposition. You ask, something collapses, and an answer walks out the other side.
iViRTH — headquartered in Ahmedabad, research team across three time zones.