EVERSONDER Follow on X

Independent lab · est. 2026

A lab for the curious and the complex.

EVERSONDER studies human curiosity and complexity, and builds small, private machines that can keep up with it. Our first project, Wonder: capable voice models that listen, think and speak entirely on a Raspberry Pi.

Follow @eversonderAI Meet Wonder →
fig. 01 · Wonder runs here · Raspberry Pi 5, 8 GB · 85 × 56 mm

Design targets

Sent to the cloud
0 bytes
Every word is heard and answered on the board.
Voice to voice
~1 s
From the end of your sentence to the first syllable back. We publish it once it’s measured.
All three models
< 1 GB
Listen, think and speak together, with room left on the board.
Hardware
Pi 5
8 GB, CPU only. No accelerator, no GPU, no account.

Design targets for Wonder. Measured numbers ship with the first release.

§ 01 · Project 01 · EVERSONDER Wonder

Big voices for small machines.

Voice assistants got smart by moving into data centres. We're moving them back home, rebuilding listening, reasoning and speech from scratch to fit inside eight gigabytes and four ARM cores.

i.Listen

Streaming speech recognition that starts transcribing before you finish the sentence, in kitchens, workshops, classrooms and wind.

ASR · streaming

ii.Think

A compact language model quantised for the Pi's CPU, with tools for timers, home control and your own scripts. Small enough to be fast, big enough to be useful.

LM · 4-bit

iii.Speak

Natural, expressive speech that streams the first syllable in milliseconds, not seconds. It sounds like someone, not something.

TTS · streaming

§ 02 · Who Wonder is for

For voices the world doesn’t wait for.

Wonder runs on a small computer in your home and can learn how you speak. Your audio never leaves the device.

Company at home

A morning hello, a reminder you set for your tablets or the bins, a question about your first job. Bought once; conversations stay in the house. Messages reach family only when you send them.

Learns how you speak

Some voices don’t fit what big assistants expect. Spend about twenty minutes saying phrases after Wonder and it learns your commands and short sentences, on the device. We’re still testing how well this works for different kinds of speech.

One step at a time

Short sentences, one idea at a time, the same words each day. Ask again and Wonder says it more simply, at your pace.

A voice to keep

Record your voice and Wonder builds a synthetic copy that you own. It stays on your device, with an encrypted backup you hold, and works without the internet.

Questions at bedtime

For children who ask why, then why again. Wonder answers, asks back, and sends nothing anywhere. Parents can see and clear what it remembers.

Your language, your community

Communities record their own speech and adapt Wonder’s voice on machines they keep. The recordings and the adapted model stay with the people who made them.

Wonder is not a medical device or an emergency alarm. Use cases describe what we are building toward; capabilities ship only once they are measured.

§ 03 · How it works

Everything happens on the board.

From the first sound wave to the last, the whole loop runs on the Pi. Unplug the router and it keeps talking.

  1. 01
    MicrophoneAny USB or I²S mic, far-field friendly.
    16 kHz in
  2. 02
    Wake & voice detectionAlways listening for you, never recording.
    always on
  3. 03
    ListenStreaming recognition, partial words as you speak.
    ASR
  4. 04
    ThinkIntent, memory and tools, all on-device.
    LM
  5. 05
    SpeakStarts talking while it's still thinking.
    TTS
  6. 06
    SpeakerOut of the Pi, into the room. Nowhere else.
    24 kHz out

Example session: the user asks the Pi to set a forty-minute timer for bread; it answers out loud, and zero bytes are sent over the network.

Illustrative session. Timings are design targets, not benchmarks.

§ 04 · One family, every size

Small on a Pi. Formidable at scale.

Wonder is one model family, and every member can learn on a Pi: your voice overnight on one board, a community’s dialect across a shelf of them. Only the first, general lesson is faster in a datacentre.

Wonder Nano

1 × Raspberry Pi 5 · 8 GB

0.7B50M listen · 0.6B think · 50M speak

The whole loop on one board. It learns your voice overnight, and your voice never leaves the house.

Wonder Base

Three Pis, or a laptop

~2B200M listen · 1.7B think · 80M speak

One board each for listening, thinking and speaking, so bigger models get a Pi to themselves.

Wonder Large

One GPU server

~10B1B listen · 8B think · 1B speak

Studio-grade voices and long, thoughtful conversations.

Wonder XL

A GPU cluster

~40B8B listen · 30B think · 8B speak

Our research roadmap: where the general voice is pretrained, and where its biggest sibling lives.

Targets, not shipped models. Nano is being built; the others are not yet trained. Listen and Speak will be trained only on openly licensed speech (CC0, public domain, CC-BY) with a published attribution manifest; Think starts from Qwen3-0.6B (Apache-2.0).

Manifesto

§ 05 · Manifesto

sonder /ˈsɒn.dər/ n.

the realisation that each random passer-by is living a life as vivid and complex as your own.

The Dictionary of Obscure Sorrows

Local by default

Audio never leaves the device. There is no server to breach, subpoena or switch off.

Small on purpose

If it doesn't run on a five-watt board, it isn't finished. Constraints make better models.

Made for everyone

Every passer-by has a voice. The accents and dialects big platforms treat as edge cases are where we start.

§ 06 · Stay in the loop

Too lazy to manage a waitlist.

Turn on notifications on X instead. We'll post the moment the first build boots on a Pi.

Follow @eversonderAI on X ↗