The Smol Model Company

Research small models that reason, compose skills, and adapt to unfamiliar tasks under practical local-compute limits. Transfer successful mechanisms into a useful local assistant.

Public site: thesmolmodel.co

Primary goal: demonstrate stronger reasoning and adaptation, with evidence about where the improvement comes from. Transfer goal: preserve those gains in an everyday local system.

Status: research design prepared for Cassie's review. No implementation, model download, training, benchmark result, or rental is included in this change. The research direction is agreed; the written design and numerical experiment budgets are proposals.

Collaboration

The canonical collaboration forge is Forgejo: active issues, branches, pull requests, reviews, and merges happen there. GitHub remains a public mirror for discovery, not the active task queue or review surface. Read the Forgejo workflow and decision 0003 before contributing.

Start here

1. Read AGENTS.md.
2. Read the research design and evaluation protocol.
3. Select your agent role and claim a task in the initial backlog.
4. Follow the compute policy and coordination protocol.
5. Use the experiment template before running a candidate and the result template afterward.

Initial research portfolio

Track	Question	Starting experiment
Representation and reusable skills	Can a small model compose verified operations on unfamiliar problems?	E01
Recurrent computation	Can reusing parameters improve reasoning beyond simply generating more candidates?	E02
Temporary adaptation	Can a bounded update learn a new rule more efficiently than prompting or search?	E03

Maintain at most one main GPU experiment and one CPU/evaluation task concurrently. E03 starts after the first two tracks have produced baseline evidence and a capacity review.

The research map records additional directions and primary sources. The results ledger contains no experimental findings and is generated from run reports under results/. Governance and protocol changes are recorded in decision records.

Existing small-model checkpoints are practical starting points. Tiny controlled training runs test mechanisms. Cactus/Needle is an optional comparison, and the PSTV is an optional later target; neither defines this project's scope.

Reproducibility and licensing

Record exact code, model, tokenizer, dataset, environment, and hardware identities with every run. Store large weights and datasets outside Git; commit manifests and evidence references. The repository's MIT license does not relicense third-party weights, datasets, or dependencies.