Roadmap
Where Run My AI is going
We’re only getting started. TokenWing shows that a model streamed from an SSD can still run fast enough for real work. From here, there’s plenty to push further: sharper predictions, smarter memory use, more models, more hardware... Come along for the ride, help us build it.
The vision
The best local AI experience
Run My AI is a simple app with the TokenWing engine inside and an autotuner that sets it up for your machine. Pick a model from the Run My AI store, hit run, and it runs.
Simple, fast, and with models bigger than your memory: the model you actually want, on the computer you already have, private and free.
The harness
Built with the community
Next to the app comes a simple harness: the part that turns a model into an assistant that gets work done. Three parts are in the plan:
- Near-lossless context compaction: already built.
- Proper context management.
- Well-engineered web search.
The rest of the harness, the agentic and coding side and how to build it, we want to shape and build with you. The direction: light and minimal, with no personality by default, and packages you add for your work, like an Unreal Engine package or a web-dev package.
Tell us what you'd want, and help us build it: harin@runmyai.org or the Discord.
Board
The work
After launch
What's coming next
The big features after launch. Ideas and help are welcome on all of them.
The harness
Light, minimal, with no personality by default, and packages for your work. Built with the community: more above.
Faster replies with drafting
The model drafts a few tokens ahead and checks them all in one step. The drafts come from its own draft head (MTP) where it has one, or from an n-gram table built from text it has already seen, like repeated code. Answers stay the model's own; replies come faster.
Adaptive memory
Uses all your free RAM to go faster, and gives it back the moment your other apps need it.
More hardware and models
macOS on Apple Silicon, Intel GPUs, and more models.
Contribute
Help with anything
Improve or add whatever you'd like: code, ideas, models, hardware tests, docs. There are no limits. If it happens to fit the vision above, that's a nice bonus. And if you're looking for somewhere to start, the roadmap has plenty.
How to start
Now and from Nov 17
- Now: join the Discord, and send your ideas to harin@runmyai.org.
- From Nov 17: the code is on GitHub, with issues and pull requests open to everyone.
Models
Model support
Hardware
Supported on day 1
- Windows and Linux
- NVIDIA and AMD GPUs, including machines with several GPUs
- CPU-only on Intel and AMD processors
- ARM processors (Snapdragon laptops, ARM Linux): builds on day 1, not yet tested on real hardware
After launch
Coming next
- macOS on Apple Silicon
- Intel GPUs