Tags

v0.5.7

v0.5.6

Rozalia 0.5.6

Curates display names for 14 models the pool serves that had none,
picked up in a survey of the live pool's connected machines. Several
were reported in a casing convention the models page had no curated
answer for, so they showed up under their raw backend ID.

v0.5.5

Rozalia 0.5.5

/v1/models says more about each model than its bare backend ID.

Every listing now names which of the caller's own machines serve a model,
and counts everybody else's by their disclosed hardware type or as
undisclosed — the same split the pool summary and account pages already
make. A model's raw ID ("qwen/qwen-3.6") is joined by a curated display
name ("Qwen 3.6"), read from a new model_names table seeded with
well-known models and editable from a new /admin/model-names page; a
migration only ever adds a name, never overwrites one an admin gave.
Where LM Studio's own API says a model's architecture and what it was
trained for, that's exposed too, as family, tool_calling, and
image_input — silent when a machine's backend says nothing, rather than
guessed.

v0.5.4

Rozalia 0.5.4

A macOS power driver arrives, and a few pages get clearer.

GPU power on Apple Silicon now reads through a new IOReport binding,
the same private Energy Model channel powermetrics, zeus-apple-silicon,
macmon, and mactop all use, wired into darwin/arm64 proxy builds
alongside the existing nvml and amdgpu drivers. CPU package power isn't
included yet: every CPU channel read a constant zero on the Mac Studio
this was tested against, so only the working GPU reading ships. A
machine's page now shows a copy-paste [backend.power] snippet built
from its own OS/arch instead of generic README prose. The account page
gains a one-click Connect button, the front page drops its AI-sounding
marketing copy for plainer sentences, and the privacy page now says how
long dead tokens and power readings are kept. Snapcraft builds no
longer pick up test-only tool dependencies that made every binary's
embedded version report as "+dirty".

v0.5.3

Rozalia 0.5.3

Power charts get easier to read, and API tokens get names you can
change.

A machine's power chart now draws a static gridline overlay behind the
data line, and gains a y-axis legend alongside a crosshair readout that
interpolates the wattage under the mouse as it moves across the chart.
The connect page is rebuilt around a generic list of integrations, so
adding a future client is a value appended to buildIntegrations rather
than a template edit, and it now points people at the pi coding agent's
rozalia extension alongside VS Code and opencode. API tokens can be
renamed after creation, alongside the existing revoke and only-mine
controls.

v0.5.2

Rozalia 0.5.2

Cross-compiled rozalia-proxy builds again for every target.

purego's dlopen bindings that the nvml power driver needs don't exist
on Windows, which broke the windows/amd64 build. Both power drivers —
nvml and the AMD GPU hwmon reader, whose sysfs path is Linux-only
anyway — now only compile into linux/amd64 and linux/arm64 builds;
macOS and Windows builds of rozalia-proxy carry no power driver.

v0.5.1

Rozalia 0.5.1

Power graphs learn to look further back, and AMD GPUs learn to report
what they draw.

A machine's power graph gains a day/week/total range toggle with
Previous/Current/Next paging, alongside an estimated energy-used figure
in watt-hours and a true bucket peak in place of a smoothed average. A
new amdgpu driver reads power straight from hwmon sysfs — no vendor
library needed — sitting beside nvml in rozalia-proxy. A machine's page
now points to the README's power setup instructions until its backend
has ever reported anything, live or historical, and machine lists link
to the dedicated Models page instead of cramming every served model
into one column.

v0.5.0

Rozalia 0.5.0

The release where machines say what they draw from the wall, not only
what they can do.

A backend can declare a power meter — NVIDIA's own telemetry through
NVML, or an external meter such as a Matter smart plug — and
rozalia-proxy reports it on every discovery tick, saying whether the
figure is a device's own estimate or what actually left the wall. The
server keeps the readings as their own time series, on their own
retention window, and a machine's page shows its live watts and
cumulative energy alongside a day's history graphed underneath; a
driver that only ever reports an instantaneous figure says so plainly
rather than showing an implausible 0.00 kWh.

/admin/config lists every setting the running server was actually
given — by flag, environment variable, or file — password included
only as whether one was set. The /v1 inference API now answers CORS
preflights, so a browser-based client on another origin can call it
directly. The landing page explains the private-hardware mode
alongside the shared pool, and the connect page points people at the
Rozalia VS Code extension before the opencode snippets. A Belarusian
translation catalog joins Polish and English.

v0.4.2

v0.4.2

v0.4.2-rc2

Rozalia 0.4.2-rc2

Second release candidate, adding unsigned macOS binaries of rozalia-proxy for Apple silicon and Intel to the release, next to the Windows installers and binaries.

v0.4.2-rc1

Rozalia 0.4.2-rc1

First release candidate with an unsigned Windows installer for rozalia-proxy, published on a GitLab Release with the bare executables and SHA256SUMS.

v0.4.1

Rozalia 0.4.1

The admin's models page names each machine's owner beside its label,
grouped by model or by machine. Labels are only unique per owner, so two
people's machines of the same name could not be told apart there before.

v0.4.0

Rozalia 0.4.0

The release where work goes to the machine that will finish it soonest,
and an account's email address is confirmed before it counts.

The server times every answer it relays and sends each request where it
would be done soonest: its prompt read at the rate the machine reads
prompts, then an answer of the usual length at the rate it writes them,
shared with what the machine is already running. A short prompt goes
where answers are written fastest, a long one where prompts are read
fastest, and a machine not yet measured is tried before it is judged. A
conversation still goes back to the machine holding its history whenever
that has room. How fast a machine reads prompts comes from a backend
that reports it, as llama.cpp does, passed on by rozalia-proxy in an
optional field that either end ignores if it does not know it, so
machines and server upgrade in any order.

/account/models lists the models a person may use, how many machines
serve each and how fast the fastest has been measured, and /admin/models
lists every model any connected machine serves with every machine behind
it and its own figures.

With a mail relay configured, the server mails a link to confirm every
account's address, and keeps an account out of all but the pages that
look after it until the link is followed; a changed address is confirmed
the same way before it becomes the login. serve takes -smtp-addr,
-smtp-username and -mail-from, with the password in
ROZALIA_SMTP_PASSWORD, and the snap maps them to smtp-addr,
smtp-username, smtp-password and mail-from. No existing account is
stamped as confirmed: once a relay is set everybody is asked,
administrators included, and account set -verified is the way back in.

A machine's stored models follow what it re-advertises, not only what it
had when it connected. The landing page names the pool in its title bar.
A tagged release asks Launchpad for an edge snap build.

v0.3.1

Rozalia 0.3.1

The release where the admin's listings sort, and the pages say what
they mean more plainly.

Every admin listing sorts by any of its columns, on the page by
choosing a heading and over the API. The admin overview shows the
server's version and the revision of the snap it came in, and draws in
yellow a snap installed from a file, which the store never refreshes,
and a version built from uncommitted changes; the machine pages do the
same for a proxy built that way.

Usage figures group their digits the way the visitor's language does,
and the landing page counts output tokens to the nearest hundred
thousand. The buttons that widen who a machine or a token reaches, let
an account lend machines, hand a machine to another owner or merge two
machine types are yellow, and a pointed-at link's words stay readable
over the blue selection fill.

Snap builds of a clean checkout are no longer stamped +dirty.

v0.3.0

Rozalia 0.3.0

The release where lenders get a say over their machines, and the pages
get a look of their own.

A drained machine now stays parked until someone returns it, across
reconnects and server restarts, and what the command line changes —
drain, revoke, delete, private — reaches a running server within
seconds. An owner may keep a machine to themselves, and a token may be
kept to its account's own machines. Owners say what their machines are
from a shared vocabulary of types, and the landing page counts the pool
by type and says how many million tokens it has processed.

Each turn of a conversation goes back to the machine that has its
prefix cached, when that machine has room. Models are named by their
context window, so qwen3.8@4k and qwen3.8@128k are two models; a bare
ID still works and goes to the widest window served.

A proxy started by hand now keeps running across snap refreshes, and
the account pages say that turning a machine on is `snap set
rozalia proxy=true`. The machine pages show which proxy version
each machine runs.

The account and admin pages are split into a page per task, and every
page is drawn in the terminal UI design language: a focused window on a
desktop, one monospaced face, and four accents for state and action.
The privacy notice sets out the right to object on its own and names
every right the GDPR asks it to.

v0.2.1

Rozalia 0.2.1

The release where the snap gets a face.

Rozalia is the mechanical horse Tytus, Romek and A'Tomek come across in
Papcio Chmiel's comic, and the snap's icon is now her head: the red
bucket, the fence of teeth, and the halo on its spring, over the green
pipe of her neck, drawn in the comic's ink and flat colour on a night blue
tile. Nothing else changes.

v0.2.0

Rozalia 0.2.0

The release where a person's data becomes theirs to see, change and take
away, and the server forgets what it no longer needs.

People can change their own name and address at /account/profile, take a
JSON copy of everything held about them from /account/export, and close
their account at /account/close. Closing erases what is about them — their
usage, sessions, tokens and unspent codes — in one transaction, and refuses
while they still own a machine or are the last administrator. An operator
can do the same for somebody who cannot sign in with `account delete`.

The server now sweeps expired sessions hourly, and usage and dead tokens
once they are past the retention window: ninety days unless
-usage-retention-days or the snap's usage-retention-days says otherwise.
Deleted rows are zeroed and the write-ahead log truncated, so an erased row
leaves the disk within the hour. Error messages name accounts and machines
by ID, never by what somebody typed.

/privacy tells people, before they sign up, what the server keeps, for how
long, who else sees a request — above all the machine that answers it —
and how to exercise their rights. The operator names themselves with
-privacy-controller and -privacy-contact, and the README gains a section
on what is left for them to decide.

Machines come back promptly after a server restart, and notice one that
went silent without closing the connection. Models served by LM Studio are
advertised with their context windows, now also at the top level of
/v1/models. A new enrollment code comes with the snap's command beside
rozalia-proxy's, and the snap's install lines point at stable.

v0.1.0

Rozalia 0.1.0

The first release: one OpenAI-compatible endpoint in front of the language
model hardware people already own.

The machines dial out. rozalia-proxy holds a WebSocket open to the server,
which pushes requests down it to whichever machine advertises the model and
has the fewest in flight, so a GPU behind NAT needs nothing opened and a
busy one makes the next request wait rather than fail.

The endpoint is /v1/models, /v1/chat/completions — streaming and not — and
/v1/embeddings, authenticated with per-account tokens, with what each
account spent recorded by model and by day.

Accounts, tokens and machines are managed from /account and /admin or from
rozalia-server: a person may be let to lend their own hardware, invited by
a link rather than by a shell on the server, and their machines drained,
revoked or retired when the hardware moves on.

It ships as the rozalia snap, with the server as a separate component so a
machine that only lends its GPU never downloads the cloud half.