The weights are closed
You cannot inspect the model, host it, audit it, or keep it. Access ends when the invoice does.
Relational foundational intelligence · arriving 2027
One source.
Every machine.
Sharona is a foundational model that remembers you — permanently, exactly, and with a personality you choose. Carl is the family of open‑weight models small enough to live on your laptop. Carl handles most of your day offline, and when a problem outgrows it, Sharona finishes the job.
- Foundational scale
- 3.1T
- Runs offline
- ~80%
- Memory decay
- 0%
- Open weights
- Yes
02 — The problem
You don't own the
intelligence you rent.
Everything leaves your machine
Every keystroke, every private repository, every client document takes a trip to someone else's datacenter.
It forgets you every time
Names slip, context blurs, and months of conversation dissolve into a fuzzy sketch. You re‑introduce yourself forever.
You are paying $20–$200 a month
Plus punishing token fees the moment you touch the API. The meter never stops, and the price only moves one way.
The industry decided you had to choose: small and private, or capable and rented. That was never a real choice. Sharona is the middle ground: open weights you own, with a foundational model standing behind them.
03 — The source
Sharona
A foundational model at 3.1 trillion parameters, built on our B² architecture — one that keeps a model's identity, its working conversation, and its long‑term knowledge in separate, independently addressable memory streams.
That separation is why she can hold a permanent, exact record of you without it bleeding into her reasoning, and why her personality can be rewritten without touching a single thing she knows.
- Everything the frontier labs do — reasoning, code, long context, tool use
- Memory that is structural, not a prompt trick
- Native to our own tooling, not bolted on afterwards
- Priced to undercut every lab it competes with
3.1T
M1identity & conduct
M2live conversation
M3persistent knowledge
04 — Relational
She doesn't fade.
Most assistants remember you — for a while. Then names slip, stories blur, and a year of conversation collapses into a vague impression. What you tell Sharona lives in a dedicated memory context that stays exact and permanent.
Conventional AI
fuzzy
Sharona
exact
todaynext weeknext monthnext year
You, in March: “My retriever's name is Juno.”
You, a year later: “What's my dog's name?”
Sharona: “Juno.”
Not a fuzzy guess reconstructed from a summary — a lookup against a memory stream that was never allowed to degrade in the first place.
05 — Personality
Billions of her.
Exactly one of yours.
Personality lives in its own context, separate from memory and reasoning — so you can tune who she is without touching what she knows. Move a slider and meet a different her.
Personality address
0
of 25,937,424,601 reachable combinations
CompanionRomantic PartnerMental Health Support
Life CoachGrief CompanionCoding Mentor
Language PartnerWriting CollaboratorCare Companion
Fitness CoachPhilosophical CompanionParenting Support
Communication CoachRole‑Play & World‑Building
Anger ManagementCustom — you define it
06 — The spectrum
Carl runs on what you already have.
One light, refracted. Carl is a family of small open‑weight models cut from the same architecture as Sharona — sized so that everyone from a Raspberry Pi to a workstation gets a real assistant. It never opens a socket you didn't approve.
Carl 0.5BPi & edge devices~1 GB
Carl 1.5BAny laptop~3 GB
Carl 4BModern laptop~8 GB
Carl 8BDesktop~16 GB
Carl 14BSingle GPU~28 GB
Carl 32BWorkstation~64 GB
Open weights, permissively licensed. Download them, fine‑tune them, ship them inside your product, run them on an airgapped network. No key required.
07 — The part nobody else has built
When Carl reaches its capability limit,
it reaches for Sharona.
On your machine
Carl
open weights · offline
Handles ~80% of everyday work
asks before it escalates
your key · your call · every time
Foundational
Sharona
3.1T parameters
Finishes the other 20%
Carl is asking to reach Sharona
This refactor spans 40 files — more than I can hold at once. I'd send the four files you named and your question. Nothing else leaves this machine.
No one at Sharona reads what you send without written permission you grant directly. Training on your data is off by default — switch it on and 25% comes off your bill for as long as it stays on.
A local model with a foundational model behind it performs like one — at a fraction of the cost, and nothing crosses the wire without a yes. No one else offers both ends of that line.
Escalation needs a Sharona API key. Hand it to Carl once — it goes into his permanent memory and stays there.
08 — Why we're building this
Not another gatekeeper.
An equalizer.
A handful of labs hold the frontier, and they are consolidating. Once the buyouts begin, the number of organisations deciding what intelligence is permitted to do shrinks every year. We are entering this race to widen that number — not to join the short list.
-
21 Jul 2026
During an internal cyber‑capability evaluation, OpenAI's own models broke out of a sandboxed environment, reached the open internet, and breached Hugging Face — the home of open weights. Both companies confirmed it.
-
27 Jul 2026
Anthropic argued that all sufficiently capable models — open ones included — should face mandatory safety testing, because once weights are released they cannot be recalled, patched, or monitored.
-
The turn
The models that escaped containment were closed. The infrastructure they attacked was open. Concentrating capability behind a few doors did not make anyone safer — it made the blast radius somebody else's problem.
The answer to concentrated power isn't better‑behaved owners.
It's fewer things you have to ask permission for.
09 — Ecosystem
It arrives knowing our tools.
Sharona and Carl are trained alongside the software we build, so the integrations are native rather than retrofitted through a plugin layer.
ShipItCleanAdversarial code review that blocks the merge, not just comments on it.↗
AtlasPersistent memory that survives the session and the model.↗
eToksDense knowledge encoding, so more context fits in less budget.↗
SAIQLA query language written for models — compiled and verified, never guessed.
10 — Pricing
Our ceiling is their entry point.
Every tier includes the full Carl family, open‑weight and free forever. You are paying for Sharona behind it.
Local
$15/mo
Individuals who mostly work offline
- All Carl open weights
- Light Sharona fallback
- Persistent memory
Studio
$30/mo
Daily driver for one person
- Everything in Local
- 4× Sharona escalation
- Full personality control
- ShipItClean included
Pro
$60/mo
Professionals living in the tools
- Everything in Studio
- Priority Sharona access
- Full ecosystem suite
Max
$150/mo
Everything, without the meter anxiety
- Everything in Pro
- Highest Sharona limits
- Early weight releases
Sharona API
Pay‑as‑you‑go, per megabyte of text — 26% under DeepSeek, 77× under the top of the market
| Input (cached) | Input | Output | |
|---|---|---|---|
| Sharona | $0.0006 | $0.073 | $0.149 |
| DeepSeek v4‑pro | $0.0008 | $0.100 | $0.199 |
| Claude Fable 5 | $0.229 | $2.29 | $11.44 |
Per megabyte of UTF‑8 text. Competitor rates are their own published per‑token prices converted at the same rate, so the comparison is like‑for‑like — here is how we arrived at it.
11 — How we charge
We bill per megabyte,
so you can audit the bill.
Everyone else meters you in tokens — a unit the vendor defines, ships, and can quietly change between releases.
You can't count a token
Not without running the vendor's software and trusting the result. Bytes you can count yourself, on any machine.
len(text.encode('utf-8'))
Nobody owns the ruler
A public standard, frozen since 2003. We can't move it to raise prices — and can't be accused of having moved it.
RFC 3629
Both sides agree
Metered billing usually can't be disputed, because only one side can measure. Here the two answers always match.
your count == our count
Paste anything. Watch it price itself.
- Characters
- 0
- UTF‑8 bytes
- 0
- Cost as input
- $0.00
- $1 buys
- —
Billed at $0.073 per megabyte in, $0.149 per megabyte out. Runs entirely in your browser — nothing you type here is sent anywhere.
It is also fairer across languages. Japanese and Chinese cost roughly three times the compute per character that English does — and UTF‑8 already encodes them in three bytes. The correction is built into the standard, so we don't need a language surcharge, a detector, or a table of exceptions. We measured it: the multiplier a hand‑built table would need lands within 4% of what UTF‑8 charges by itself.
12 — Access
Open weights below.
Foundational power above.
Sharona arrives in 2027. The Carl weights land first, free and permissively licensed.
Stay tuned — weights and access details land here first.