Eco AI: the greenest token is the one we never generate.

Training and serving large models consumes electricity, water for cooling, and rare hardware — and the biggest systems run around the clock across warehouses of accelerators. “Green AI” emerged as a research idea in response: measure the cost of results, not just the results, and prefer approaches that get most of the value for a small fraction of the energy.

For a product rather than a paper, we think eco AI comes down to four commitments:

• Small enough to matter. Use the most compact model that does the job, on hardware that sips rather than gulps.
• Do not generate what nobody asked for. Long, repetitive replies are the cheapest waste to eliminate.
• Do not run two machines when one is idle half the day. Schedule background work into the gaps.
• Say only what you can measure. A confident carbon number with no meter behind it is marketing, not ecology.

Conway-Retrain 12B — our current model, retrained on top of a Gemma 4 base — is served through Apple’s MLX framework on a single 16GB iMac. That machine handles live chat, memory, and the background learning loop. There is no fleet of GPUs waiting for traffic and no second region idling for redundancy.

Apple Silicon is the quiet hero here: unified memory lets a 12-billion-parameter model run on a desktop that draws household power, and it does so without a fan you can hear from across the room. The trade is honest — frontier systems will beat it on breadth and trivia. Its edge is the self-learning loop, not raw scale.

The model stays replaceable

The checkpoint sits behind an interface, so when a more efficient model appears it can be swapped in without touching memory, safety rules, or your ownership of what you taught. The earlier 188M Conway-Omega checkpoint — small enough to be almost free to run — is deprecated in-product but remains open-source on Hugging Face.

Every token costs compute, so the cheapest energy saving in the whole stack is not generating tokens nobody wanted. conw.ai enforces short-answer token ceilings so replies stop when they have answered, and repetition guards catch the model when it starts circling and cut it off.

The same instinct runs through the learning loop. Only safe, useful, short replies are eligible to teach the model; down-rated and contaminated examples never enter the queue, so compute is never spent training on noise.

A self-learning AI could easily double its footprint by training constantly on a second box. conw.ai does not. Eligible examples wait in a guarded queue. Training begins only when there is enough clean material and serving is idle — and it yields whenever chat traffic needs the machine. Live chat first, learning second.

The output is a candidate checkpoint that must load and pass its checks before anything live changes, which means no cycles are spent promoting a model that would be rolled back an hour later. You can watch what that process kept and refused at conw.ai/learning.

Qualitative on purpose — we will not put numbers in a table we cannot stand behind.

Sending every question across the internet to a warehouse and back is a cost in itself. Conway Entity removes it: the same self-learning loop — local model, local memory, weight-level retraining — running entirely on your own machine, with nothing leaving it. No network hop, no shared cluster, no data centre in the loop at all. It is in early access for Pro and Max plans.