krummholzWorking forests.
Living waters.
A planned domain assistant with dynamic weights for forestry and fisheries: make source material easier to use, interpret structured records and explain results from reviewed analytical tools.
Release target: less than 1% hallucination on a defined evaluation, alongside quality, privacy and security checks. This target is not yet achieved or a guarantee across all uses.
Run it your way.
Krummholz models are released with open weights under Apache 2.0, with dynamic, user-reviewed learning in development. Hosted inference and training on your private data will be paid services. Third-party components retain their applicable licenses and notices. Atlantic v8 (27B), the model Katahdin will be distilled from, is out in beta with open weights. Katahdin is still in development and not yet available.
The domain comes first.
The base is chosen.
Katahdin is our desktop-focused forestry and fisheries direction. It will be distilled from Atlantic, our 27B general-purpose model, the same teacher as Acadia. The teacher is now Atlantic v8, released in beta on November 15, 2026. Katahdin will inherit the same written constitution and the same safety harness. We first planned to build Katahdin on a 35-billion-parameter mixture-of-experts core; that core is deprecated. Katahdin is not trained yet and has no release date.
Atlantic’s training already includes answers grounded in Maine forest inventory records, flags for bad rows in forestry data, and summaries from a point-cloud tool. Katahdin-specific training and domain validation are the next milestones.
Atlantic v8 model card → · Atlantic v8 weights on Hugging Face → · How v8 was trained →
Make information
easier to work with.
We will select one or two concrete workflows with domain review to guide Katahdin-specific development and evaluation. Candidate tasks include interpreting inventory fields and units, locating relevant source documentation, organizing observations and explaining a validated analysis.
Longer-term areas of interest include growth, mortality, migration and climate-related analysis, classification, and domain coding assistance. These require suitable data, methods and expert validation; they are not current forecasting capabilities.
From a field code to an explanation.
A practitioner asks what a measurement code means. Katahdin would locate the approved field dictionary, explain the definition and units, identify missing context and link the answer to its source.
This intended workflow can be evaluated independently of whether a model can produce a plausible ecological prediction.
Language explains.
Reviewed methods calculate.
Katahdin’s planned adaptation will prioritize modern authoritative open sources, permitted curated records and carefully reviewed synthetic examples. Historical texts can support data-pipeline experiments but should not establish current operational guidance.
Numerical calculations, unit conversions, missing-data handling and statistical uncertainty belong in validated tools. The model’s role is to help frame the question, explain the returned result and preserve the important qualifications.
Our public forestry and fishing datasets are preliminary starting points with explicit sampling and privacy limitations.
What we will test
- Correct use of source definitions, units and dates.
- Faithful explanations of reviewed analytical outputs.
- Recognition of missing data and unsupported conclusions.
- Separation of measured observations from generated examples.
- Domain review of the selected pilot workflows.
Useful where
the work happens.
Katahdin is planned for desktop and hosted workflows with permissioned tools. Approved reference material may be available locally, but the distilled model’s size and task routing are not settled. Actual compatibility, memory use, network requirements and offline limits will be reported after testing.
Support judgment.
Keep the evidence visible.
The initial pilot will not claim validated ecological forecasting, statewide population estimates, harvest prescriptions or real-robot control. It will support bounded information tasks and reviewed analyses, with source references and escalation when evidence is insufficient.
MODEL PERSONALIZATION / IN DEVELOPMENT · BASE IN BETA
Learning with a review step.
Our planned “What I learned today” review lets users accept, reject or edit proposed learning. Local and hosted models may diverge. Preserving learning through updates remains a goal to test.
Planned safeguards include source controls, bounded updates, adversarial testing and rollback. These features are not released. Learning will not grant new tool permissions.
How dynamic weights will work →IN DEVELOPMENT · BASE IN BETA
The early launch release is out.
Katahdin has no date yet.
Atlantic v8 and Atlantic Lite v8.1 were released with open weights on November 15, 2026 and are still in beta. Katahdin is in development: it is not trained yet and has no release date. It will be distilled from Atlantic v8 and will inherit the same constitution and harness. We plan to release open weights after evaluation and release review, with paid hosted inference and private-data training. Dynamic weights personalization remains in development. The company’s full launch is planned for January 2, 2027.