Operating system for desktop and robots · 64-bit ARM · version 0.15.3 · Build 629
A robot needs a system that is never late.
SOFTrobots OS is built entirely on its own – from the loader to the database, in Embedded Swift, without Linux. With Develop you write a program while the system checks it at every key; for the real-time part it proves that every period is long enough. And the AI runs right on the device – even models larger than its memory.
Develop and real time, recorded in SOFTrobotsVM on 14 Nov 2026 – every time is real (2:24 min, with sound).
Measured in SOFTrobotsVM on a Mac with Apple M3 Pro. Start: median of three runs on 15 Nov 2026, 4 cores, 8 GB (on a Mac mini M4: 0.57 s); Develop and the real-time proof: 14 Nov 2026; AI: 15 Nov 2026, 4 cores, 8 GB. Real robot hardware has not been measured yet.
Experimental. Never put important data into it. Everything can change – formats, API and data can be lost with an update. Never use it for real robots, machines or vehicles.
Develop
A few lines.
A finished window.
Develop is the workshop inside the system. You write in Reflex, a language that reads like a recipe – on the right the window appears, with buttons, slider and chart. Develop checks at every key and suggests the fix right away when something is wrong.
1count = 0 2step = 1 3 4page "Counter" 5 value count label "count" 6 button "Add" 7 slider step from 1 to 10 8 chart count 9 10when "Add" is pressed 11 count += step
A shortened version of the template “Counter” (the template also has a heading, a text, Reset and a test). The window on the right is drawn after it; the real one is in the picture further down.
Measured in SOFTrobotsVM (7 cores, 8 GB) on 14 Nov 2026 while filming the Develop film, with the templates “Counter” and “Wrist hold”; Develop shows these times itself, anew at every check.
Forms and lists
A text field, a button and a list that is still there after a restart – an excerpt of the template “Notes”. The list is a table in the program's own folder.
Charts in one line
chart draws a value over time – the counter, a measurement or the angle of a joint.
12 templates to start
Empty, Notes, Counter, Stopwatch, Tea timer, Shopping list, Quiz, Arm buttons, Arm move, Camera watch, Clap to stop, Melody – ⌘N, pick one, Enter.


Real time
Not hoped for.
Proved.
An arm that reacts late is dangerous. So for every real-time part Develop works out how long one period takes in the worst case – instruction by instruction. Only what fits safely into 80 % of the period may go onto the real-time core.
1target = 40° 2periods = 0 3 4realtime every 1 ms 5 joint = robot.state(joint: 5) 6 if periods % 10 == 0 7 arm.move joint 5 to (joint.angle 8 + clampStep(error: target - joint.angle)) in 10 ms 9 periods += 1
If the time cannot be proved, Develop shows it at once not proved – and “Save as Program” does not put such a script onto the real-time core.
Template “Wrist hold”, proof in Develop on 14 Nov 2026; the proof counts for a 1 GHz core, with a safety margin.
Develop checks at every key – the real-time part too. The proof is a card beside the code.
Develop compiles to ARM64 machine code, signs it with this robot's key and puts it onto the real-time core – with fixed memory, without a heap.
Period and budget are in the signed manifest. If the contract does not fit the core, the program does not start.
A click on E-STOP or ⌘+Esc: the stop goes over CAN to the motor – in the VM 10–12 µs from the emergency stop in the kernel to the stop frame sent. The flight recorder keeps the last seconds.

AI models on the device
Every model runs.
With an explanation.
A small model fits into memory and answers smoothly. A large one doesn't – and runs anyway: SOFTrobots OS fetches from the disk only the parts it needs for the next word. That makes it slower, but it works. All on the device, no cloud.
SOFTrobotsVM on a Mac with Apple M3 Pro (36 GB). Qwen3 4B: 4 cores, 8 GB, median of three runs on 15 Nov 2026. Kolibri-1 (Q4_K_M): 8 CPUs, 18 GiB, disk image on an external SSD over USB, 18 Nov 2026.
What is a model?
A file full of numbers
A language model is a large file holding billions of numbers, learned from an enormous amount of text. Qwen3 4B is 2.4 GB, Kolibri-1 47.5 GB.
One word at a time
From your question the model works out the next word, appends it and works again. For every word it has to read the numbers it needs. Strictly speaking these are pieces of words; we simply call them “words”.
Why memory matters
When the numbers sit in main memory, reading them is fast. When they first have to come from the disk, the model waits for the disk. So the speed depends on how much has to come from the disk for each word.
Small is fast. Large works too.
SOFTrobots OS doesn't turn away a model that is too big for memory – it loads it anyway.
Small: fits entirely
The whole model sits in memory. Every word is read from memory alone.
Large: the experts come from the disk
A fixed part stays in memory. The rest is a cache for experts; whatever is missing, the system reads from the disk.
Bars not to scale. Values: SOFTrobotsVM on the Mac with Apple M3 Pro; Qwen3 4B with 4 cores and 8 GB (15 Nov 2026), Kolibri-1 with 8 CPUs and 18 GiB (18 Nov 2026); the expert cache held 8.7 GB.
What are experts?
Many new large models are split into small sub-networks called experts (“mixture of experts”, MoE for short). For each word a router picks only the ones that fit; the rest stay idle. Kolibri-1 has 384 experts in every layer and uses 6 of them. So for each word it reads only about a twentieth of the file – and that is what lets a model larger than memory run at all.
An illustration, not a measurement: which experts sit in the cache changes with every conversation.
Whatever every word needs – attention, vectors, the output head and the shared expert – is always ready.
In every layer it picks the 6 best-fitting of 384 experts for the word. One Kolibri-1 expert is only 2.6 MB.
Experts that are needed often stay in memory. Only what is missing comes from the disk.
The larger the cache, the more often an expert is already there. Nothing in the system needs to be reconfigured for that.
How fast? Measured with Kolibri-1.
The same AI engine, once inside SOFTrobotsVM and once as a test tool running directly on the Mac. The numbers show what more memory for the cache achieves.
Words per second. Mac with Apple M3 Pro, 36 GB, on 18 Nov 2026; Kolibri-1 Q4_K_M (47.5 GB) on an external SSD over USB (10 Gb/s), 32 words per run. VM: 8 CPUs.
Which model for what?
Honestly measured. Green means smooth enough to read along; amber means worth it when the answer can wait.
| Kind | Example | Speed | Good for |
|---|---|---|---|
| Small, fits into memory | Qwen3 4B Instruct, 2.4 GB | 31 words/sVM, 4 cores, 8 GB | everyday Ask, short questions, explanations |
| For code, fits into memory | Qwen2.5 Coder 7B | 19 words/sVM, 4 cores, 8 GB | questions about programs |
| Medium, fits with 16 GB | Qwen3 14B, 9.0 GB | 12 words/sVM, 7 cores, 16 GB | longer, more precise answers |
| Large with experts, bigger than memory | Kolibri-1, 47.5 GB, 6 of 384 experts | 1.0–1.1 words/sVM, 8 CPUs, 18 GiB · 5.68 directly on the Mac with a 16 GB cache | good German, answers that can wait |
| Large with large experts | Llama 4 Scout, 65.4 GB, one expert 81 MB | 0.3–0.7 words/sdirectly on the Mac, 16–24 GB cache | runs and answers correctly; every missing expert means a lot of reading |
Qwen values: SOFTrobotsVM on the Mac with Apple M3 Pro, median of three runs on 15 Nov 2026 (Qwen3 14B on 12 Nov 2026). Kolibri-1 and Llama 4 Scout: 18 Nov 2026, models on an external SSD over USB. All models 4-bit (Q4_K_M). Models without experts (“dense” models) also run when they are bigger than memory – but they read the whole file for every word and therefore stay slow. With little memory, a model with experts is the right choice.
The system tells you
Settings › AI shows for every model whether it fits. Ask puts it in the window title and shows the real speed while it answers. The system deliberately shows no estimated time – it would only be a guess.

This line is the real text from Ask.
Private and open
The AI service reads only the /models folder, has no network and writes nothing. Models come in the formats you can get them in anywhere – swap them without any extra tools.
Language: GGUF
One file into /models/language. Today the engine knows the architectures Qwen2, Qwen3, Qwen3 with experts, Qwen3 VL (images), Llama, Llama 4 and Kolibri-1.
Detection: ONNX
A detection model goes into /models/detection. Camera reads it directly and turns it into 8 bits while loading – no converting on the Mac. If the engine doesn't know a building block, it says so plainly.
Not signed, but checked
Models are not programs, so they are not signed. The reader checks every size against the file; to prove it, small models were deliberately damaged and read in 20,000 times. A broken file does not bring down the AI service.
This holds for the AI service, Ask and Camera. The exception: Develop's assistant (Explain, Ask Develop) asks a service on the internet – only with a key of your own; the preview has none.

At the desk
Calculate, plan, print.
All on the device.
SOFTrobots OS is also a desktop for every day: spreadsheets, letters, contacts and appointments in open formats, printing over the home network, folders on your Mac and 3D models to look at. No account, no cloud – the files stay on the device.
Calc and Write
Calc works out sheets with formulas and templates such as the household book; a formula can also be described in a sentence. An .xlsx stays an .xlsx – nothing is converted silently. Write writes letters as .odt and PDF.
Contacts and Calendar
People and firms with birthdays, groups and attached files; vCard 2.1 to 4.0. Appointments are typed in one line ("Dentist Fri 14:30"), and Buddy reminds you.
Printing
⌘P in Write, Calc, Reader and Image: the system shows the print dialog with a preview and only what the printer can do. Network printers over IPP, as PDF. Programs never see your printers; the printer sees only "SOFTrobots document".
Folders on your Mac
Files opens folders of your Mac over SSH/SFTP – always encrypted, with post-quantum key exchange too. You compare the Mac's fingerprint once; if it changes, Buddy barks. The robot only connects outwards; nothing new listens on it.
Parts: 3D models
Parts shows 3D models to turn, measure and cut, with sizes, volume and whether the model is watertight. Files shows a preview. The file is read by a service of its own with no network and no file rights.
Files
Your favourite places on the left, folders you cannot write with a lock, cut and paste, dragging between two windows, ZIP with a password and the time slider that brings back older versions.




Security
Secrets in a safe.
Visible, but sealed.
A key you cannot see is hard to understand. So SOFTrobots OS hides nothing: secret files stand in Files as rows of steel – only no program can open, copy or restore them.
The safe
The robot's own key, API keys, the server's keys and guest list, passwords for network folders and Web's cookies are sealed. Only the program a secret belongs to and the kernel read it – not even Console or the time slider.
Locks that explain themselves
Every locked place says who may change things there and why, with a button that goes straight there. Places that are only read are shown as red glass by the system – no program can take that off.
Checked models
Only Settings › AI and the model download change AI models. Every model file gets a SHA-256 check; a changed file is refused. Models are not signed – they stay files in standard formats.
Rights for the system only
Important rights – all files, offering AI, installing models – go to system programs only. Programs you install yourself stay in their own folder.
Encrypted from the start
The data lies on the disk with AES-256-GCM per block. The key comes from your password (Argon2id); the password itself is never stored.
Signed
Every program is signed and runs in its own address space; updates are signed packages. An update that does not start is taken back by itself after three tries.


The system
All our own.
From the loader to the database.
Kernel, drivers, file system, database, network, desktop and programs – in Embedded Swift. C only in the EFI loader and in the JavaScript cage, assembly only where nothing else works. The only third-party code: the JavaScript engine QuickJS-ng, locked in a cage with six system calls.
Web, Image, Camera and Video – recorded in SOFTrobotsVM on 15 Nov 2026, every time is real (1:58 min, with sound). Web shows this website; photos CC0 from Wikimedia Commons.
Rights from the manifest
Every program is signed, runs in its own address space and gets only the rights in its manifest. A program sees a file only when it is given to it.
SOFTrobotsFS and the database
Copy-on-write, checksums on everything, AES-256-GCM per block. After a power cut the old or the new state holds – never anything in between. The database lives in the file system.
Buddy, the watchdog
Reads the logs, barks at faults and bites programs that take more memory than they announced.
Network only encrypted
IPv6 first, only TLS 1.3, names only over DNS over HTTPS, time over NTS. A port without a listener does not answer at all.
Programs
Develop, Ask, Camera, Video, Image, Files with a time slider, Write, Calc, Contacts, Calendar, Parts, Web (JavaScript in a cage), Reader, Sound, Database, Benchmark, Activity, Server, Console, Settings – and games such as chess.
Small
The whole system: a download of 53 MB, 181 MB on disk after the first start (an 8 GB image with holes) – with the desktop, all programs and Develop's compiler.


Architecture
How the system is built.
Swift from bottom to top.
Five layers, one main language: kernel, services and programs are Embedded Swift. C only in the EFI loader and in the JavaScript cage, assembly only where Swift cannot reach – at start, for exceptions, for the thread switch, at the system registers and for fast crypto.
Languages in our own code
- Swift173,78591.3 %kernel, drivers, services, programs; plus Mac tools and tests
- Assembly3,2041.7 %kernel 1,374 · programs 1,830
- Shape language (.shape)2,2671.2 %one definition for API, rings, tables and network
- JavaScript1,4580.8 %Web's page mirror for scripts in the cage
- C1,3740.7 %EFI loader 573 · JavaScript cage 801
- C headers9350.5 %BootInfo, EFI, ABI structures
- Reflex7560.4 %templates and programs for Develop
- Python and Shell6,4803.4 %Mac tools only – never run in the system
- Total own code190,259100 %lines including comments
Third-party code, counted separately
QuickJS-ng – 82,889 lines (MIT licence), Web's JavaScript engine. Runs only in the cage, not in the total.
Counted on 14 Nov 2026 with wc -l; technical datasheet, section 4.
File formats
Open formats –
no licence fees.
Every format is our own Swift code. The decoders run in the program's address space, never in the kernel – a malicious file can hit only that one program. And we only build formats that anyone may use without paying licence or patent fees.
Pictures
PNG (W3C/ISO) was built to be patent-free; for baseline JPEG the patent claims have expired, for GIF the LZW patents expired in 2003/2004. WebP comes with Google's royalty-free licence, SVG is a royalty-free W3C standard, BMP and TIFF are simple, documented formats.
Documents
PDF is ISO 32000 and free to implement. Text and Markdown are plain UTF-8 text. OpenDocument is an open ISO standard (ISO/IEC 26300), and so is .xlsx (Office Open XML, ISO/IEC 29500).
Contacts, appointments, 3D
vCard (RFC 6350) and iCalendar (RFC 5545) are open IETF standards. STL, OBJ and PLY are simple, openly described formats; 3MF is an open specification of the 3MF Consortium.
Video
VP9 comes with Google's royalty-free licence (BSD code, patent grant); Matroska and WebM are open containers. MJPEG is JPEG pictures – that is how Camera stores its recordings.
Sound
Opus is IETF RFC 6716 and royalty-free, FLAC is open and free, WAV and PCM are simple, documented formats. Opus is read in WebM and Matroska.
Archives
Deflate (RFC 1951) is free to use – one own inflate for PNG, ZIP, PDF, TIFF and Web.
Data
CSV and JSON are open, simple text formats. SQL is read-only, in Console and Database.
AI models
GGUF is the open format of llama.cpp (MIT licence), ONNX an open standard of the Linux Foundation. We read both directly – no converting, no third-party runtime.
Our own formats
Our own formats – described in the technical datasheet.
R = read, W = write
Deliberately not built: H.264 and H.265/HEVC. Patent pools charge licence fees per device for them – the decision is: never H.264.
Variants, limits and speed of every format, checked against the code: technical datasheet, section 9. For the WebP reader the algorithm was ported from libwebp to Swift (BSD licence, in the file header).
How it is built
Developed with the help of AI.
Reviewed and tested by the developers.
Since its first start on 28 October 2026: one change at a time, measured before and after – what brings nothing is removed again. Every failure is written down (180 entries, F1 to F180), every decision recorded, and before a version 31 device tests and 581 tests on the Mac run.
Try the preview
SOFTrobots OS
on your Mac.
An app and the finished system, for Macs with Apple silicon. Unpack, double-click, set up the robot – and write your first program in Develop.
Put SOFTrobots OS.app anywhere, Applications for example – the system is inside it.
Signed with the Developer ID of SOFTrobots GbR and notarised by Apple.
A name and a password – the password is never stored.
Needs: a Mac with Apple silicon (M1 or newer), macOS 15 or newer, 4 GB of free memory, about 2 GB of free space.
Contact
Curious?
Write to us: info@softrobots.de. Every number with its source and conditions is in the technical docs, the programming interface in the API.