Abstract reasoning
Visual puzzles where the rule must be inferred from a few demonstrations and applied to a new case.
In active development
Eunoia Labs · Research preview
Eunoia Omega
A compact reasoning model for problems it has never seen before.
Scroll ↓
Eunoia Omega is a compact reasoning model developed by Eunoia Labs, the AGI research division of SHV Groups. It is built for one purpose: solving problems it has never encountered before.
Instead of relying on knowledge absorbed during training, Omega learns the underlying rules of a new task from just a handful of examples and applies them with precision.
Fewer than 500 million parameters. On purpose.
Omega is a fraction of the size of today's frontier models, and that is intentional. We believe genuine reasoning comes from architectural discipline, not scale.
A model that can't memorise its way to an answer has to understand its way there. Omega's size is the constraint that forces it to think.
- parameters
- <500M
- parameters
- kinds of reasoning
- 2
- kinds of reasoning
- ARC-AGI benchmarks
- 3
- ARC-AGI benchmarks
It doesn't jump to an answer.
When Omega meets a new problem, it first builds its own understanding of what it is looking at, then works its way to a solution.
- 01
Understand
Identify the elements involved, how they relate, and what changes between examples.
- 02
Hypothesise
Form candidate explanations of the hidden rule that could account for every example.
- 03
Test
Check each hypothesis against the evidence and discard the ones that don't hold.
- 04
Commit
Answer only with a solution that survives the test.
The result is reasoning that is deliberate, grounded and checkable, rather than a single best guess.
Two kinds of challenge. One ability underneath.
Interactive reasoning
Unfamiliar environments where Omega has to explore, work out how the world behaves and pursue a goal without being told how.
Together, these test the core of general intelligence: the ability to adapt to the new.
Reasoning that should get cheaper the more of it you do.
A system with frozen knowledge pays the same price to rediscover how an environment works on its fortieth attempt as on its first. People don't learn that way.
Our central claim is that Omega's cost per environment should fall as it encounters more of them. If it stays flat, the claim is unsupported, and we will publish it as unsupported.
Measured where scale alone doesn't win.
We evaluate Omega on the ARC-AGI benchmark family, designed to measure generalisation to novel problems rather than recall. They are among the few benchmarks where scale alone doesn't win, which makes them the right test for what Omega is built to do.
ARC-AGI-1
Infer a rule from a few grid examples and apply it to a new one.
Result pending verification
ARC-AGI-2
A harder successor built to resist memorisation and brute-force search.
Result pending verification
ARC-AGI-3
Interactive environments explored without instructions.
Result pending verification
How results will be reported
- Results are published on official evaluation sets once verified, with conditions stated plainly.
- No comparison is drawn from a single run.
- Every reported score carries seed-level dispersion, not a point estimate.
- The protocol was fixed in advance and is not changed mid-programme.
What we are not claiming
- No trained results have been published. Nothing on this page is a statement about achieved performance.
- A negative result on our central claim is a publishable outcome, and will be reported as one.
What we publish, and what we keep.
Released
- Evaluation protocol and its version history
- Benchmark results, including negative ones
- Training curves, hardware and compute
- Ablations and error analysis
Retained
- Data and how it is generated
- Representation and model design
- Training methods
- Source code
This boundary was fixed in advance.
In active development.
Omega is not yet publicly available. We'll share access plans alongside our first published results.
Model details
- Developer
- Eunoia Labs, SHV Groups
- Model type
- Compact reasoning model
- Parameters
- Fewer than 500 million
- Focus
- Abstract and interactive reasoning
- Evaluation
- ARC-AGI-1, ARC-AGI-2, ARC-AGI-3
- Status
- In development
- Availability
- Not yet public