Nova Nova NovaOS NovaOS
CODING LANG PROJECT · 31 OCT – 5 NOV 2026

Nova roadmap: from a new language to an AI running on its own operating system

In six days, this project went from "can we create our own coding language" to a language with safety built in, a compiler, and a whole operating system written in that language. That system now runs DeepSeek, a 500 GB AI model, and writes answers faster than the best-known program for the same job.

98pull requests merged
179automatic tests passing
5hardware drivers written in Nova
4.6words a second from DeepSeek (started at 0.12)

NovaOS vs llama.cpp, at real speed

Words written at each program's speed on the same question. This is a replay of those results, not a live feed.

NovaOS 4.6 words a second
0
llama.cpp 3.41 words a second
0

NovaOS writes about 1.35 times as many words in the same time.

About 1.35xllama.cpp's writing speed on a short, warm question
About 37xthe speed of the first try, which ran on a desktop PC
Same answerschecked after each speed-up, so speed didn't cost quality

Goals reached

  • First answer on the bare server 4 Nov
  • 1 word a second 4 Nov
  • 2 words a second 4 Nov
  • 3 words a second, and past llama.cpp 5 Nov

Still getting faster

  • Reading long questions faster, where llama.cpp still leads
  • Checking answers word by word against llama.cpp
  • Timing long chats of 2,000 and 8,000 tokens

How fast DeepSeek writes on NovaOS

Words written per second. A "word" here is a token, the small chunk of text the model works in. Longer bars are better. Each speed-up was checked to make sure the answers stayed the same.

The early runs include reading parts of the model from disk. From "chat mode" on, the numbers are for a repeat question with the model already loaded. The grey bars are llama.cpp, the well-known program for running AI models on a computer's main processor, on a second server with the same chips and memory. Its copy of the model stores some parts in a different 8-bit format and its disks differ, so this is a close comparison, not an exact one. Both figures count only the words written. llama.cpp's was measured (3.41, then 3.42 on a repeat). NovaOS's is calculated from one whole-answer run, because that run didn't log writing time separately; the next runs will measure it directly.

The road so far

31 OCT

The Nova language

Nova started as a small language run by an interpreter (a program that reads Nova code and does what it says). Safety was chosen as its defining feature, and it was built in six steps:

  • Step 1, no surprise changes: values are fixed unless you say they can change, and numbers can't silently overflow.
  • Step 2, type checking: mistakes like adding text to a number are caught before the program runs.
  • Step 3, errors you can't ignore: anything that can fail has to be handled.
  • Step 4, permissions: a program can only touch the files, websites and settings you allow.
  • Step 5, untrusted input and secrets: outside text must be checked before use, and passwords can't leak into printouts.
  • Step 6, packages: add-on code gets only the permissions you give it, with an audit report.

Nova was also made easy for AI helpers to write, and got its first game, a text adventure called Lighthouse Keeper.

#1 Nova for AI agents #3 Lighthouse Keeper
31 OCT – 1 NOV

Security tools and safe web apps

To help find holes in web apps, Nova gained a security toolkit and a web server that is safe by default.

  • Secret scanner that spots more than 20 kinds of leaked keys, and a website safety grade from A+ to F.
  • Checks for known-bad add-ons, a tamper-proof security log, and a one-command nova security menu.
  • Web server with safe logins, cookies and protection against fake form posts.
  • Built-in database that can't be tricked by injected commands, plus a firewall that blocks probing.
  • Safe page templates, access rules, two-step login, encryption, safe file uploads and backups.
  • Key files like .env and .ssh are always blocked, and risky actions can ask a human first.
#2 #4 #5 #6 #7 #8 #9 #10 #11
1 NOV

The compiler (step 7)

nova build turns Nova code into a real Linux program, written straight to machine code with no outside tools. This is what made an operating system possible. Later updates made compiled code about twice as fast.

#12 compiler #14 #22 2x faster
1 NOV

NovaOS is born

The rule: everything in NovaOS is written in Nova, except a tiny start-up piece that can't be. In one day it went from first boot to a usable small system.

  • Boots on a PC, saves files to a disk, and has a clock that stops runaway apps.
  • Apps must ask before touching your files, and are type-checked before they run.
  • Memory is cleaned up automatically, and apps can be installed while it runs.
  • A full-screen text editor with cut, copy, paste, find and replace.
  • Installed apps run as machine code for speed.
First boot: NovaOS starts up in a virtual PC.
#13 first boot #15 #16 #17 #18 #21 #23 editor #32 machine code #19–#33
1 – 2 NOV

AI models running inside NovaOS

Step by step, NovaOS learned to run AI language models, from a tiny story writer up to DeepSeek V4.1 Flash.

  • Chat with a local AI (Ollama) from Nova.
  • Number lists for AI maths, memory above 4 GB, and every processor core working together.
  • Reads standard model files (GGUF and safetensors). SmolLM2 and Qwen answer questions.
  • "Experts" models, where only a few parts of the model work on each word.
  • Fast SSD driver (NVMe), faster maths with AVX2, and support for two processor chips.
  • DeepSeek's layers, its lookup tables and its experts read from disk when they don't all fit in memory.
  • Tests that try to break the model checks on purpose, to prove the checks really work.
DeepSeek runs on a desktop PC: about 8 seconds a word.
#34 Ollama #38 all cores #41 SmolLM2 #43 experts #44 NVMe #49 DeepSeek #55 full model #35–#64
3 – 4 NOV

Real hardware: a physical server

With a physical server to test on, NovaOS got the drivers it needed to run on real machines with no Linux underneath.

  • Ready for big servers (up to 255 cores) and newer UEFI start-up.
  • Drivers written in Nova: USB keyboard, SATA disks, the server's disk controller, and a network card with ping.
  • A small read-only status web page served by NovaOS itself.
  • Never writes over another system's disk.
  • Chat mode keeps the model loaded between questions.
  • Hands-off test runs: a bare-metal test can boot NovaOS and return to the normal system on its own.
3 Nov, 23:47: NovaOS boots on the bare server. 4 Nov, 02:50: first DeepSeek answer on bare metal.
#66 big server #69 UEFI #70 USB #71 SATA #72 network #73 web page #81 disk controller #84 chat #89 hands-off tests
3 – 5 NOV
Target beaten

The speed race: faster than llama.cpp

DeepSeek on NovaOS went from about 8 seconds a word to under a quarter of a second, with identical answers the whole way.

  • Attention (the step that looks back over earlier words) shared across all cores.
  • Both processor chips used properly, each working on the memory next to it: 0.65 to 0.44 seconds a word.
  • The busiest loops rewritten as machine code, including the memory cleaner: down to 0.34.
  • Table maths tuned for this server's processor, and smarter sharing of work between cores.
  • Less wasted work: cached tables, disk reads done together, skipping steps that change nothing: 0.217.
  • --remember N lets chats run longer than 2,048 tokens.
5 Nov: about 4.6 words a second, about 1.35 times faster than llama.cpp (3.41) at writing the answer to the same warm question.
#77 #86 #87 #90 both chips #93 #94 #95 #96 past llama.cpp #97 --remember
NEXT

What comes next

  • built, being timed Reading questions faster. Long questions are where llama.cpp is still well ahead. NovaOS now reads question words 32 at a time; the server is timing it against one at a time.
  • in progress Checking answers against llama.cpp. Both programs score the same "bicycle" answer word by word and compare their top 5 guesses, to prove NovaOS's maths is right.
  • in progress Long chats. Timing a 2,000-token chat, then an 8,000-token one with --remember.
  • planned Driver for the server's own network card, so NovaOS can talk over the network on bare metal.
  • idea Processor speed setting. A BIOS setting may be holding the server's processors back; changing it could add speed for free.
  • idea Web server, next stage: handling many visitors at once.

Known limit: NovaOS can't use a graphics card, so all the AI work runs on the main processors.