Pre-alpha · nothing released yet

A foundation lab.

The long-term goal is training models from the ground up. The starting point is deliberately narrower: build the evaluation we trust before we train anything we care about.

The plan

Foothills are the ground you cross before the summit. The name is the plan.

  1. 01

    Benchmarks

    Build evaluation harnesses we trust before we train anything we care about. A lab that cannot measure cannot improve.

  2. 02

    Fine-tuning small models

    Learn the full training loop — data, tokenizer, training, eval, release — at a scale where iteration is cheap and mistakes are recoverable.

  3. 03

    Build up from there

    Each rung earns the next. Scale is a consequence of competence, not a substitute for it.

Models

Elevation tracks capability tier. Continent tracks regional and language specialization.

Planned models, their mountains, elevations and roles. Nothing is released.
Glyph Model Continent Elevation Role Released
Everest Asia 8,849 m Flagship — the largest model we train
Aconcagua South America 6,961 m Spanish-language model
Denali North America 6,190 m Reserved
Kilimanjaro Africa 5,895 m Fast tier — the one anyone can walk up
Elbrus Europe 5,642 m Reserved
Vinson Antarctica 4,892 m Reserved
Kosciuszko Oceania 2,228 m Reserved — the smallest of the seven

Each glyph is contoured from the mountain's elevation at a constant 1,500 m interval, so ring count is the tier by construction. Siwalik — the literal foothills of the Himalaya — is the codename series for experimental and pre-release checkpoints, and the house mark.

Code

Each released artifact gets its own repository. Apache-2.0, and nothing is published to a registry yet.

Roadmap

Where the work actually is, rather than where it is going.

Phase Focus State
0 Repo, branding, project foundations In progress
1 Benchmark harness and first eval suite Next
2 Fine-tuning pipeline for small open models Planned
3 First published model and eval results Planned
4 Pretraining experiments Horizon

Principles

Measure before you train

Every model claim ships with the eval that backs it, and the harness that produced it.

Small and honest beats large and vague

A 1B model with a reproducible number is worth more than a big claim with none.

Reproducibility is a feature

Seeds, configs, and data provenance are part of the result, not an appendix to it.

Publish the negative results too

The failures are most of the signal.

If something here is useful to you, or wrong, open an issue.