The Goldenfish G1 API is now available.
Turn context into structured decisions. Free for one week with inline choices. No API key required.
aspaslabs.com/docs/
we opened the Goldenfish G1 API today and made it FREE for the next 7 days
if you wanted to try the model but didn't feel like downloading weights, setting up serving and dealing with the rest of that stuff, now you can just hit the API
G1 is our 576M decision model, so instead of generating text it takes a set of options and chooses between them, which has been useful for things like tool selection, routing and other cases where you just need a model to make a decision and move on
the API is still pretty new, so part of the reason we're opening it like this is because we actually want people to use it, stress it a bit and tell us what sucks
docs.aspaslabs.com
Introducing Goldenfish G1.
Our first model for typed decisions. 576M parameters, built to select tools, route requests, and rank candidates without generating text.
Open weights available now.
aspaslabs.com/goldenfish-g1/
Introducing Goldenfish G1.
Our first model for typed decisions. 576M parameters, built to select tools, route requests, and rank candidates without generating text.
Open weights available now.
aspaslabs.com/goldenfish-g1/
G1 reaches 81.00% on our 14,802-example development suite.
On Open Jev Snake validation, it scored 51.85% hard accuracy (407/785; Choice ties excluded). Unfamiliar low-option state-to-candidate binding remains a limitation.

ALT Goldenfish G1 results and limitations on a darkened abstract painted background. Large orange pixel numerals show 81.00% on the internal development suite of 14,802 examples and 51.85% hard accuracy on Open Jev Snake validation, with 407 correct out of 785 one-hot rows after Choice ties are excluded. These are different evaluation surfaces, not directly comparable. The internal development result is not an external generalization score. Unfamiliar low-option state-to-candidate binding remains a limitation. Julia 1 was much stronger in the same Snake harness.
Goldenfish G1 weights are available now on Hugging Face:
huggingface.co/aspaslabs/gol…
The API is not live yet. We're working to make it available as quickly as possible.
Read the release note: aspaslabs.com/goldenfish-g1/
been heads down on the system one model at @aspaslabs for a while, we're close to shipping it
meanwhile we've already started writing the training plan for an LLM from scratch
now comes the annoying part finding enough money and compute to actually run the thing
cc @horizzon3507
Soon, Aspas Labs will release its first System One model, the result of extensive work on training, evaluation, and inference.
We're also preparing to train an LLM from scratch, with the training plan being developed as we seek the funding and compute to make it happen.
Aspas Labs is now in the world.
We're an independent AI lab from Brazil. Our first work on Goldenfish and the Intelligence Network is already up.
Thanks to @itxtoledo for the domain, and @semfence for soon powering our training and experiments.
aspaslabs.com








