Your Own AI v0.6.0
Private AI on your own machine - offline by default, online only when you choose.
Your AIs know you better, run bigger models on the machine you already
have, and choose between them by what actually runs well on it.
Highlights
- Your AIs know you as a whole, not as a list. From what they have
learned, they keep a short summary of who you are - "How your AIs see
you" - written on your device and rewritten as things change. You can
read it any time, and if it is wrong you fix the fact it came from and
it rewrites itself. Every fact is yours to edit or forget, traceable to
the conversation it came from, and none of it leaves your machine. See
Memory. - Bigger models on the machine you already have. Mixture-of-experts
models split between your graphics card and main memory - measured 4.5x
faster on an 8 GB card - and the catalog now has a mixture-of-experts
for every size of machine, from a 4.8 GB file for 12 GB laptops to
753B for workstations. See Big mixture-of-experts models and New
models. - Routing that measures, not assumes. On your device, the app times
how fast each model actually runs on this computer as you use it, and
picks by those numbers - not a spec sheet. Online models follow your
preferences instead. See Routing. - It fails politely. The first-run recommendation sizes to your whole
machine, a load that cannot succeed is refused with a plain sentence,
and the app never crash-loops on a model that took it down. See The
app fails politely. - Models live where you say. Store models on any drive, with free
space checked before every download. See Model storage. - Records that tell the truth. A conversation is in your records the
moment you send; a stopped reply is kept and marked. See Your
records. - Apple Silicon: MLX engine, as a preview. Optional, per model, chats
only, no speed promise. See Apple Silicon.
A damaged model file can't take the app down
- A model file that is incomplete or corrupted (a browser download that
stopped early, a copy that went wrong) could crash the app at startup
while it looked through your models folder. Every model header is now
read within hard bounds; such a file is listed as "Damaged file" with a
delete button, is never loaded or offered, and downloading the same
model again replaces it. - A finished download is checked to read as a model before it gets its
final name.
Big mixture-of-experts models run on small graphics cards
- A mixture-of-experts model bigger than your graphics memory now runs
with its rarely-used parts in main memory and the rest on the card,
instead of being refused or crawling. Measured on an 8 GB RTX 4060 Ti
with 32 GB of RAM: Qwen3.6 35B-A3B went from 6 to 26 tokens per second,
gpt-oss 20B from 14 to 23, and the 35B loads in seconds instead of most
of a minute. The gate is main memory: a 32 GB machine qualifies for the
35B, a 16 GB machine is told so honestly. The split is sized from the
model file itself - as many of the model's expert layers as fit stay on
the card, the rest go to main memory - so a bigger card keeps more.
Model cards say "Runs here - split with main memory", and "Best for
this computer" can recommend these models. - Downloaded models show this computer's measured generation speed
("~27 tok/s measured") once you have used them - your number, not a
benchmark's. - A reply's tokens-per-second now comes from the local engine's own
timing, so a short one-line answer no longer reads as tens of
thousands of tokens per second. - Models that always think before answering (LFM2.5) now answer directly
in chat and think in Report mode, so a short question never ends inside
the model's reasoning with nothing visible. gpt-oss replies no longer
stop at the end of its reasoning: one of the app's generic stop markers
was the same token gpt-oss uses to close its reasoning channel, which
cut every reply off before the answer. gpt-oss is no longer used for
memory extraction, which it cannot format.
New models - mixture-of-experts for every size of machine
-
Two workstation-class additions for very large machines: DeepSeek V4
Flash (284B mixture-of-experts, 13B active, 1M context; 128 GB+ of
main memory) and GLM-5.2 (753B mixture-of-experts; 384 GB+). Both
ship as multi-part files: the app downloads the parts in order (each
part resumes on its own), shows "part 2 of 3", treats the set as one
model, and keeps them off machines that cannot run them. -
Every model file's layout is now checked against its size when it is
read, so a download cut off inside the weights - not just in the header- shows as damaged instead of failing to load.
-
Where the maker ships a speed-up file for a model (Nemotron 3.5
Lightning's multi-token head, DeepSeek V4 Flash's DSpark draft), the
app downloads it after the model, keeps it beside it, and the engine
uses it to answer faster - several drafted tokens checked per pass,
which matters most when a model's experts live in main memory. A model
downloaded earlier offers "Get its speed-up file" on the Models page. -
The split between graphics card and main memory learns from your
machine: after a model loads, the app measures what the card actually
held and uses that for the next load, so it settles on the most the
card can carry. If a split ever runs out of graphics memory, the app
retries at once with the experts in main memory instead of reporting
the model too large. -
Four additions to the offline catalog, all mixture-of-experts and all
running on the new split-memory path: LFM2.5 8B-A1B (Liquid AI -
a 4.8 GB file for 12-16 GB machines, the first MoE that fits them),
Granite 4.0 Tiny (IBM - 3.9 GB, a 1M-token context, Apache-2.0),
Nemotron 3.5 Lightning 30B-A3B (NVIDIA's August 2026 reasoning
model, for 32 GB machines) and Ling-mini 2.0 (Ant Group - 16B-A1.4B,
MIT, for 24 GB machines). LFM2.5 asks you to accept Liquid AI's license
once before its first download.
Apple Silicon: MLX engine (preview)
- Macs with Apple Silicon can add an optional MLX engine (Settings >
Engines, about 50 MB, checked against a pinned release before it runs)
and then fetch MLX versions of supported models - 15 in the catalog,
from Ministral 3B to Qwen 3.6 35B - with "Get the MLX version" on a
downloaded model's row, and an MLX chip on catalog cards once the engine
is installed. Chats run on MLX once a model's MLX version is installed;
project work, images, memory and everything else stay on the standard
engine, and your existing model file stays. Whether MLX is faster
depends on your Mac - the app makes no promise, and nothing changes
unless you install it. Remove on the engine card puts everything back. - MLX downloads resume file by file, show live progress, and pick back up
after navigating away or restarting the app.
Model storage
- Choose where models are stored - and not just chat models: vision files,
speed-up drafts, engine packs, OCR models, and the Build agent all live in
the same folder and move together. The control is on the offline models
page and in Settings > Components. It shows how much is stored and how
much room the drive has, with a Change button that moves everything to a
folder you pick - a second drive, an external SSD. Downloads follow, moves
roll back safely if interrupted, and if the drive is missing later the app
falls back to its default folder until it returns. - Downloads check disk space up front and say plainly what is missing
("needs 22.4 GB free on the models drive, 9.1 GB available") instead of
failing mid-file with "no space left".
Downloads
-
A resumed download now proves it is continuing the same file it started
(If-Range): if the model was re-uploaded upstream in between, the download
starts over instead of stitching two versions into one broken file. -
A model file the engine rejects ("layout doesn't match") is reported in
those words with the fix (delete and re-download), sits out the session,
and no longer counts against the graphics card - two rejections of one
bad file could previously switch the whole app off CUDA. -
If CUDA was ever switched off by the safety ladder, the notice offering
"Try GPU again" now stays available instead of appearing only once. -
Several models can download at once, each card with its own progress
and button. Starting a second download used to make the first card look
idle. -
Coming back to the Models page shows a download's real progress right
away, and a download the app was closed in the middle of picks itself
back up from where it stopped. -
A vision model is two files; the card now says which one is in flight
("Downloading model · 1 of 2", then "Downloading vision support · 2 of
2"), shows downloaded of total, and the Download button states both
sizes up front ("3.2 GB + 0.9 GB vision") - the bar no longer looks
like it started over.
Routing
-
The GPU + RAM split now checks how much of a mixture-of-experts model
actually fires on every token. Catalog models activate a few percent -
their experts are cold mass and the split is fast. A converted
"surgery" model that activates most of its width gets the honest
"Runs slower" grade instead of a promise the hardware can't keep
(it still runs if you pick it). -
Project work: local picks for agent sessions rank tool capability first (the speed lean only decides between equals); a local model counts as ready for agent work only when it fits fast and, once measured here, keeps 8+ tokens a second - the same bar for side-work on this device and for "keep whole project sessions on this device". When a session goes online by default while an equally capable local model is ready, the Model button says so.
-
The Auto modes now judge models by how they actually run here, not by
where their weights sit: a mixture-of-experts model running GPU + RAM
counts as running well (it no longer loses to a small model just for
using main memory), and a dense model that does not fit the graphics
card is honestly "too large" instead of "runs slower". -
"Prefer fastest" ranks by each model's measured speed on this computer
once you have used it; a model that is slow to load here must be
clearly better before it replaces the one already loaded; and a
question bigger than a model's reading room goes to a model that can
hold it. -
Settings > Routing shows your models as routing sees them - how each
runs, measured speed, load time, reading room, capability scores and
project readiness - next to the recent decisions. Nothing new to set.
Memory
- Memory is now two things, not four: each AI has ONE memory (entries you've
given it, and moments it's learned with you), and what every AI knows
about you lives in ONE place - Your Memory. Project sessions' notes are
now called project notes, so "memory" always means those two. - Your AIs now keep a short summary of who you are - a few sentences
written on your device from your remembered facts and notes, shown at
the top of Your Memory ("How your AIs see you") and given to every AI
alongside the facts. It rewrites itself in the background as your memory
changes - including updating things that have since happened - so if it
is ever wrong, fix or forget the fact it came from and it follows.
Duplicate learned facts tidy themselves up along the way. Nothing
leaves your device, and pausing memory pauses this too.
Reading room follows the turn
- A long attachment or a long conversation no longer fails with a raw
"exceeds the available context" error. Before sending, the app checks
whether the running model's context can hold the turn; if this
computer can afford more, it reloads the same model with more room
first - you see the usual loading hint - and then sends. If the
server still finds the turn too big, the app grows to the exact count
and sends again by itself. - When nothing this machine can afford would hold it, the app says so in
plain words - how much room the turn needs, how much the model is
running with - and offers what the AI's mode allows: an
online-and-offline AI can send it to an online model that holds it (an
explicit choice for that turn, since an attachment would leave your
device); an offline-only AI is never sent online quietly - the offer is
to switch it to Auto - Online and Offline first; a pinned model gets
"let Auto choose here". Health questions stay on your device whatever
the size, and the online offer only appears when your plan includes
online models. - In Auto - Online and Offline, a turn too long for every model on your
device routes to an online model that can hold it by itself (long
conversations; attachments only with consent), and the reply says why. - Models can now run with up to 131,072 tokens of context where the
hardware carries it (the ceiling used to be 32,768 for every model, no
matter the card or the model's own limit) - always sized against your
graphics memory and clamped to what the model was trained for. - The context cost of hybrid models (the Qwen 3.5 family, Nemotron 3.5,
Granite 4, LFM2.5) is now counted from the layers that actually keep
attention, not every layer - they were being charged several times
their real cost, which kept them at smaller contexts than the card
could carry and under-graded their fit. - The attachment meter counts with the running model's own tokenizer and shows its real context, with the room the app can make for a longer turn - chips go amber when a reload will hold them and red only when nothing on this machine can.
The app fails politely
- The first-run recommendation now sizes to your whole machine, not just
the graphics card: a model that fits the card but not the system's
memory is no longer offered. A big card in a machine with little memory
used to be handed a model it could not load - and the app went down
silently, every time. - A model load that cannot succeed is refused with a plain sentence
("Not enough free memory to load this model safely right now - 1.4 GB
free, about 2.7 GB needed. Close some other apps and try again, or pick
a smaller model.") instead of the app vanishing. - If the app ever stops mid-load, it will not automatically retry the
same model on the next start - the crash loop is gone. Picking the
model yourself still gets a fresh attempt. - Mixture-of-experts models are labeled as such in the catalog (Ornith
1.5's 35B is now "35B-A3B (MoE)"), which is also what unlocks their
GPU + RAM split on cards they do not fit.
Conversations
- The conversations list is ordered by last activity: continue a chat or
a project from days ago and it comes back to the top, with the time of
its last turn on the row (hover shows when it started). With more than
one AI, a row of their avatars at the top filters the list to one AI.
Projects
- The Projects tab (project notes shared by all your AIs) loads in a
fraction of the time on installs with a long history: every AI's
records are read at once instead of one after another, the last list
shows instantly when you come back, and the "warming up" message only
appears while records are genuinely still warming after launch. - Project memory is now called project notes everywhere, and the Projects
tab explains what projects are when you have none yet. - A read of your records that stalls now gives up after a minute with a
clear error instead of hanging the page.
Code in replies
- Short code blocks (up to 14 lines) now stay right in the reply where they
were written. Longer code is lifted into the code panel, and its first
lines show in place, fading into a "Show all N lines" chip - so a reply
can never read as if its code went missing - The app's helper programs start with their console windows created
hidden: no flash, and no Windows Terminal error dialog on machines where
Windows Terminal is the default terminal. Helper programs left behind by
an earlier launch that did not finish are cleaned up more thoroughly at
the next start.
Your records
-
A conversation exists in your records from the moment you send your
first message: it appears in the conversations list while the reply is
still being written, and your words are kept even if the reply is
stopped, fails, or the app closes mid-answer. (Nothing used to be
written until the reply finished.) -
Stopping a reply keeps what was written so far on screen, marked
"Stopped here", and records it as far as it got, flagged as stopped -
so your records show the conversation as it actually happened. -
History stays reachable even when older storage generations are switched
off - reading your records no longer depends on every generation being
awake at once. -
Replies from online models now record the provider's own fingerprint
for the backend that answered (when the provider sends one), alongside
the model name: the provider's claim, inside your encrypted record.
Offline replies already carry the hash of the exact model file.
Other
-
Loading a model now shows its proper catalog name ("Qwen 3.6 35B-A3B
(MoE)", not a cleaned-up file name), and when a load takes more than a
few seconds, short lines explain what the wait is - the model moving
into memory, once - and that it stays loaded for the questions after.
Fast loads show none of it. -
"Your AIs can now go online" shows once, the moment a plan activates.
-
Removed the disabled records-tidying command.
-
The Models page's picks for this computer rank by where a model would
run first - on the card, then split with main memory, then processor -
the same order the first-run recommendation uses, so both screens name
the same model for the same machine. -
The offline models page carries the same header controls as every
other page, Conversations included.
Elsewhere
- Every model, offline and online, now has its own page at
yourownai.net/models - with a check of what runs on your machine, the
same sizing the app uses.
Downloads
See assets below for platform-specific installers. Verify any download against SHA256SUMS.txt.