Ommny Research / San Francisco and Marseille The sensorium issue, No. 02

Sight. Sound. Touch. Taste. Smell. and language, one model.

Ommny builds omnibus reference models that interpret the world and all five senses as data. One perception model encodes sight, sound, touch, taste, smell, and language into a single token space, trained on the omnibus corpus.

Request API access Read the research

Portrait of a woman half lit, half in shadow, dramatizing how little of any scene a single channel carries — Cover *The sensorium, held.* One channel of six, exposed. Photograph from Pexels.

Perception API: v1.2
Corpus tokens: 2.4T
Shared context: 128k
Median latency: 380ms

01 Sight 02 Sound 03 Touch 04 Taste 05 Smell 06 Language

01 / The argument

Five senses, one model.

Why we stopped building six kinds of model and trained one that treats the world, and every word about it, as a single stream of data.

The history of machine perception is a history of separate rooms. Vision models in one, audio in another, touch reduced to a regression problem, taste and smell politely ignored. Each room developed its own vocabulary, its own benchmarks, its own idea of what an object is. A strawberry has lived a divided life: a redness score here, a spectrogram of crunch there, a molecule string somewhere else entirely.

Ommny was founded on a refusal of that division. Our omnibus corpus captures the world the way witnesses do, with every channel at once, and our reference models are trained to place a retinal frame, a sound pressure wave, a force curve, a volatile compound, and the word strawberry into one shared token space¹. Not six models wired together at the seams. One model, one geometry.

A word for warmth and the measurement of warmth belong to the same geometry. That is the whole thesis.

The consequence is not merely neat engineering. When channels share a geometry, they correct one another. A shadows-only video gains structure from the sound of footsteps on gravel². A robotic gripper learns fragility from the acoustics of a crunch. Language stops describing the world from a distance and starts living inside it, one embedding among neighbors it can touch.

Macro photograph of a woven beige textile where individual fibers are visible — *Plate 02.* Woven linen at 40x. In the omnibus corpus this weave is stored beside the words coarse and warm, beside the sound it makes pulled taut, and beside the pressure it returns under a thumb.

Scale matters as much as unity. The omnibus corpus now spans 2.4 trillion tokens across six channels³, versioned and audited like code, so a dataset can be cited the way a paper is⁴. Perception at this scale stops being pattern matching and becomes reference: a model you interrogate about the world the way you interrogate a library. Ask it what rain smells like, and it answers with petrichor, with the ozone beforehand, and with the particular hiss of rain on hot asphalt, because all three live in the same space.

1. The omnibus corpus spans licensed and first-party sensory captures: 2.4 trillion tokens across sight, sound, touch, taste, smell, and language, each record carrying provenance metadata.
2. Crossmodal anchoring is described in Ommny Research Report ORR-26-02, section 3, with ablations across anchor density schedules.
3. Snapshot of 2026-09-20. Counts exclude duplicate retakes captured within 48 hours at the same station.
4. The provenance ledger is available to audit partners under the Corpus Charter, including consent records for every human participant.

The spread

The senses, in sequence.

Five channels of the omnibus corpus, photographed, captioned, and encoded. Leaf through the spread, then meet the sixth channel, language, which describes all the rest.

Scroll to leaf through

Extreme close-up of a hazel eye, the iris resolved into fine radial texture

Sight

A retinal frame and its caption enter the same stream. A sunset and the word dusk become neighbors in the space.

Channel 01 / visual / 1.1T tokens

White wave foam breaking on a dark sea seen from above

Sound

Pressure over time. The model hears the room in the reverb and the weather in the white noise.

Channel 02 / acoustic / 0.6T tokens

Macro photograph of the ridges and lines of a human palm

Touch

Force curves, slip, texture under load. Grippers learn fragility the way readers learn subtext.

Channel 03 / haptic / 0.3T tokens

Mounds of colorful spices arranged in trays at an outdoor market

Taste

A spectrum, not a sentence. Sweet, sour, salt, bitter, umami, and every mixture in between.

Channel 04 / gustatory / 0.1T tokens

Macro photograph of an orange daylily dusted with pollen

Smell

Volatile compounds, ten thousand of them nameless. The corpus gives them geometry before words.

Channel 05 / olfactory / 0.2T tokens

+ Language,
the sixth channel.

The one that describes the other five. In the O-1 token space it is not a summary of perception, it is perception, one embedding among neighbors.

Read the research

02 / The instruments

A small family of large references.

Every model descends from the corpus, not from each other. S-1 and E-1 are distilled from O-1 with anchoring intact, so a robot running E-1 offline still reasons with the same geometry as the flagship.

The Ommny model family
	Model	Parameters	Channels	Context	Availability
01	O-1 Omnibus Flagship reference	480B, 37B active	6 of 6	128k	API, waitlist
02	S-1 Sense Perception encoder	30B dense	6 of 6	32k	API, standard
03	E-1 Edge On-device perception	3B quantized	4 of 6	8k	Containers, offline
04	C-1 Corpus The dataset engine	Not a model	6 of 6	Versioned	Enterprise

Distilled, not fine-tuned. Benchmarks and method: ORR-26-04.

03 / The record

Research, in public.

Every finding is published as an Ommny Research Report, numbered, citable, and preregistered before the runs. Hover an entry to see its plate.

The full index

04 / In the field

Where perception is the product.

Macro photograph of palm lines, standing in for tactile sensing

Plate 03. robotics, as the model receives it.

Robotics

Grippers that know ripe from ready. O-1 fuses camera frames with force telemetry and motor audio, so a pick arm feels a bruise forming before it sees one.

Field note: greenhouse harvest trial, 99.2 percent undamaged

Read ORR-25-11

Foam breaking over dark water, an acoustic scene rendered visible

Plate 04. accessibility, as the model receives it.

Accessibility

The world, narrated as it happens. Live sight and sound become spoken description with a 380 ms median, turning crossings, counters, and doorways into one continuous sentence.

Field note: street-crossing trial, Lyon, 380 ms median

Read ORR-25-02

Woven textile macro, the kind of surface E-1 inspects at line speed

Plate 05. industrial inspection, as the model receives it.

Industrial inspection

Fabric, welds, circuit boards. E-1 runs on the line at sixty frames a second and catches the defect whose signature is half visual, half acoustic, and entirely missable by either alone.

Field note: loom trial, defect recall 0.994

Read ORR-26-02

A daylily heavy with pollen, the visual cousin of a scent cue

Plate 06. immersive media, as the model receives it.

Immersive media

Films you can lean into. Editors attach scent, texture, and low-frequency cues to the timeline, and audiences feel the scene arrive a beat before it cuts.

Field note: scent-synced screening, 12 compounds, 400 attendees

Read ORR-25-06

05 / Admissions

Request API access.

The perception API is provisioned like a laboratory instrument, deliberately. The sensory dataset engine is licensed the same way.

01
Apply

Tell us what you need to perceive. Applications are read by researchers, not routers.
02
Scope

Together we define the channels, the context window, and the guardrails for your deployment.
03
Build

Keys in days, not quarters. Perception with provenance, from the very first call.

Tiers

Research Free, rate-limited, citation expected
Scale Volume pricing, dedicated anchors
Enterprise On-prem E-1, corpus licensing

Request access Read the research

We reply within five working days. Every key ships with the Corpus Charter attached.