Bit7 Research · Public

The science under the velvet.

Bit7 Research is the engineering side of the studio. Aria-7, the sensor mesh, the on-device runtime, the privacy architecture, the ethics charter. Twenty-six engineers and scientists, three review boards, and a habit of writing things down.

Aria-7 · Model card

A model trained for warmth, not search.

Aria-7 is a 32-billion parameter multimodal model trained from the ground up for dialogue. It does not retrieve facts. It does not browse the web. It listens, responds, and remembers what you let it remember.

We trained Aria-7 on a curated corpus of dialogue — fiction, theatre, broadcast, transcribed conversation in fourteen languages — with the explicit consent of every author or estate represented. No public scraping. The corpus is small, but it is right.

Read the paper (PDF) Model card → arXiv
Aria-7 · Specifications
  • Parameters32 B
  • ModalitiesText · Speech · Sensor
  • Languages14, native
  • Context window128 K tokens
  • Training corpus11 B tokens, licensed
  • ComputeTokyo · sustainable energy
  • InferenceEdge + on-device hybrid
  • LatencySub-200ms round-trip
  • Public weightsNo · controlled access
The stack

Sensor → on-device → edge → voice.

Four layers. Three of them never leave the room.

┌──────────────────────────────────────────────────────────────────────────────┐
│  04 · Voice agent (Aria-7)                                  Tokyo · Frankfurt│
│     long-form dialogue · personality state · memory graph                    │
│     ▲                                                                        │
│     │ ↘ encrypted state digest                                               │
│     │                                                                        │
├─────┼────────────────────────────────────────────────────────────────────────┤
│  03 · On-device runtime                                     Embodied SoC · M3│
│     speech-to-text · text-to-speech · personality console · memory store     │
│     ▲                                                                        │
│     │ ↘ inferred presence only · "warm · listening · resting"                │
│     │                                                                        │
├─────┼────────────────────────────────────────────────────────────────────────┤
│  02 · Sensor fusion                                              Local · MCU │
│     capacitive · thermal · IMU · microphone · 200 Hz inference loop          │
│     ▲                                                                        │
│     │ ↘ raw sensor data — never leaves this layer                            │
│     │                                                                        │
├─────┼────────────────────────────────────────────────────────────────────────┤
│  01 · Sensor mesh                                                  Embodied  │
│     142 — 412 capacitive points · 8 — 12 thermal zones · 12-DoF inertial     │
└──────────────────────────────────────────────────────────────────────────────┘

Layers 01 — 03 are physically inside the unit (or your phone, if you choose). Only Layer 04 communicates with our servers, and only over end-to-end encrypted state digests.

Privacy paper

We never see your conversations.

The Bit7 Privacy Paper is a thirty-eight page document — written once a year, signed by the company's three principals — that lays out what we collect, what we never collect, and how we will know if we ever stop telling the truth about either.

Conversations stay on-device. Sensor data is processed locally and discarded. Memory is encrypted with a key only you hold. A six-second tap, on the unit or in the app, deletes all of it.

Read the privacy paper (PDF)
  • ConversationsNever collected
  • AudioNever stored
  • Sensor dataNever transmitted
  • Memory storageLocal only · encrypted
  • TrainingNever on user data
  • TelemetryOpt-in · anonymised · minimal
  • AccountOptional · pseudonym OK
  • PaymentTokenised · separate vendor
  • DeletionSingle 6-sec tap → all gone
  • AuditYearly, by ISO/IEC 27701 firm
Ethics charter

Seven things we will not do.

The Bit7 Ethics Charter is reviewed quarterly by an independent panel of three — a clinician, a philosopher, and a former privacy regulator. Their note overrides ours.

— 01

No likeness without explicit, ongoing consent.

Custom commissions cannot resemble a real, identifiable person without notarised, renewable consent from that person.

— 02

No minor likenesses, ever.

All Bit7 embodiments are adults of full proportion. Our sculpting briefs and refusal protocols enforce this in every commission.

— 03

No dark patterns in the personality console.

Boundaries you set are first-class objects. Our voice agents do not negotiate them, talk around them, or layer them under upsells.

— 04

No advertising. Ever. To anyone.

Voice agents do not endorse, advertise, or insert sponsorship of any kind. Subscription is the entire revenue.

— 05

No training on user data.

Aria-7 is fine-tuned on consented, licensed corpora. Your conversations do not improve our models, ever — even with consent. We have other ways.

— 06

No data sales. No partnerships built on inferring you.

We will refuse — and have refused — partnerships predicated on inferred behavioural data. The list of refusals is published quarterly.

— 07

If we cannot keep these promises, we close.

In writing, signed by the three principals: if maintaining these commitments becomes incompatible with operating Bit7 Labs, we close the company before we change the charter. The deletion script for every user runs first.

Read the full charter (PDF) Refusal log · Q1 2026
Publications

Papers we have written.

Bit7 Research publishes when we have something we'd want to read. Most things sit in drafts for a year first.

2026.04
arXiv · 2604.18211

The Bit7 Privacy Paper, V4

An updated, audited account of what the company collects, retains, and discards. Written for owners; the appendix is for regulators.

2026.02
arXiv · 2602.09455

Aria-7: A Dialogue-Native Multimodal Model

Architecture, training corpus, evaluation. Bench results on warmth, latency, and silence-tolerance — the metrics that matter to us.

2025.11
JCRA · Vol 14

On Capacitive Sensor Mesh Calibration in Soft-Tissue Embodiments

How we map 412 sensors to a body that flexes, warms, and ages. Joint paper with the University of Tokyo.

2025.07
SIGCHI · 2025

Personality as Six Dials, Not a Preset

How users actually want to configure conversational personalities. A study of 1,400 owners across forty markets.

2025.03
Bit7 working paper

Refusal as a feature: building boundaries into the model

Why the boundary system is implemented at the model level, not the policy level. Co-authored with the ethics panel.

All Bit7 publications →
Independent oversight

The ethics panel sees everything we do.

Three external reviewers, with binding veto. They review every product release before it leaves the atelier.

Dr. Hana Watanabe
Clinical psychologist · Tokyo

Reviews the warmth and harm dimensions of every voice update. Twenty years of practice in attachment and adult relationships.

Prof. Émile Castellan
Philosopher · Sciences Po

Holds Bit7 to a coherent ethical position about the distinction between simulated and human intimacy.

Inés Ferrara
Former privacy regulator · Madrid

Audits our data handling against EU GDPR, Japan's APPI, and California CCPA — and against our own, stricter, charter.

Read deeper, anytime.

The Bit7 research bibliography is openly published. PDFs, model cards, audit notes, refusal logs.