AIRUNCODE 1.5.4 IS HERE
Something pretty important changed in AIRUNCODE.
The model is becoming interchangeable.
Claude, GPT, other providers, an API endpoint, Chinese models, or a model running locally on your own machine can now operate inside the same environment and get access to the same AIRUNCODE tools.
That means you don't need a frontier model burning expensive tokens for every stupid little step of a workflow.
You can bring in a frontier model when you actually need that level of intelligence, let it solve the hard part, and then have a local model continue working with what was already done.
The expensive model doesn't need to be the worker carrying bricks all day just because it was the architect for five minutes.
That's the direction I've wanted from the beginning.
And we've pushed the graphical side much further too. The models can now work with the graphical engine in a way that lets them build, recompile, inspect, change and rebuild.
Think LEGO.
Instead of giving the AI a finished LEGO set and letting it move a few pieces around, we're giving it the pieces and the tools required to keep rebuilding the thing until the result is actually useful.
So a "sample" can finally be a real sample: something that can be compiled, changed and compiled again as many times as necessary until it becomes a clean product.
The new subscription system is live now too.
I've already given one AIRUNCODE subscription away and I'm giving another one away. More importantly, I want people actually running this workflow now.
Use a frontier model. Use your endpoint. Use a local model. Mix them.
And if you break AIRUNCODE, tell me.
That's literally useful to me right now. π
Replies
The "expensive model is the architect for five minutes, not the guy carrying bricks all day" line is what sold me. Letting a local model take over after the hard part is done sounds like a real way to save money..
@david_turner12Β That's exactly the idea. The expensive model should spend its time on the part that actually requires expensive intelligence
Once the architecture, plan or difficult reasoning is done, having Claude or GPT carry every brick afterward just feels wasteful. I'm trying to make that handoff between frontier and local models feel completely normal instead of being a whole new workflow
Which local models have worked best for you so far? And can a normal laptop run one without being painfully slow?
@oliver_graf1Β Yep. A normal laptop can absolutely run useful local models β the trick is not expecting a small local model to suddenly become Claude or GPT
That's actually why I'm building the workflow this way. Give the frontier model the nasty reasoning problem, then let a smaller local model continue with the context and handle the work it can actually do well.
Hardware and quantization make a huge difference, though. I'm doing more testing around this now because I want to publish actual AIRUNCODE benchmarks instead of throwing random βbest modelβ names at people.
What hardware are you running? CPU/GPU and RAM/VRAM? I can probably give you a much more useful answer from that.
A new step for AIRUNCODE.
I'm closing the loop between desktop and mobile.
AIRUNCODE's agents run on your computer, but you shouldn't need to stay in front of that computer to follow their work, approve actions, or continue a workflow.
I'm now preparing AIRUNCODE's mobile experience as a proper Android application.
The idea is simple:
Your desktop does the heavy work.
Your phone keeps you connected to it.
The Android/TWA infrastructure is now being integrated, including verified app-to-domain communication and the necessary separation between web subscriptions and the Google Play environment.
I'm not calling this a Google Play launch yet. That announcement will come when the application is actually available there.
This is another step toward what I want AIRUNCODE to become: not another interface around a single AI model, but an environment where your models, tools, projects and workflows remain connected regardless of which device you're using.
Desktop is where the work runs.
Mobile is how you stay in control.