krummholzBetter support.
A human route forward.
A planned customer-service model with dynamic weights for answers grounded in approved policies, useful request classification and dependable escalation when a person needs to take over.
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 Acadia will be distilled from, is out in beta with open weights. Acadia is still in development and not yet available.
NOVEMBER 9 / BUILT ON ATLANTIC
One teacher.
A customer-service model.
Acadia will be distilled from Atlantic, our 27B general-purpose model: trained to match Atlantic’s answers on customer-service work. The teacher is now Atlantic v8, released in beta on November 15, 2026. Acadia will inherit the same written constitution and the same safety harness. We first planned to build Acadia on a 35-billion-parameter mixture-of-experts core; that core is deprecated. Acadia is not trained yet and has no release date.
Atlantic v8’s GGUF files are about 17 GB (Q4_K_M) or about 19 GB (Q5_K_M), plus a vision projector, for workstations and consumer graphics cards. File sizes are not RAM requirements. Atlantic’s training already includes request classification on a sample of the BANKING77 dataset. Our earlier 4B Acadia candidate has become Atlantic Lite, our model for phones.
What comes next
We still need to test the quantized builds on actual devices, including memory, speed and heat. A compact file is a useful starting point, not proof that the complete product runs well on every workstation.
Atlantic v8 model card → · Atlantic v8 weights on Hugging Face → · How v8 was trained →
Understand the request.
Find the right next step.
The first Acadia pilot will cover a single, representative text-based customer-service workflow. It is intended to identify what a customer needs, retrieve the relevant approved policy and provide a supported answer or a clear handoff.
Account-specific decisions, refunds and other consequential actions require the customer’s authorization rules and appropriate tools. A plausible model response is not permission to act.
A delivery question, with context.
A customer asks where an order is. Acadia would classify the request, use only authorized order information, explain the applicable policy and escalate missing or conflicting information to the designated service team.
This describes an intended workflow, not a current demonstration or live integration.
Teach the task.
Keep the policy current.
Atlantic gives Acadia a starting point that already handles tool calls and request classification. The size of the distilled model and its release checkpoint still depend on evaluation. Curated service examples include permitted intent data and separately labeled synthetic scenarios. Synthetic answers will need approved policy support; a generated policy is not a valid training target.
Customer policies and other changing information belong in a controlled retrieval collection. Customer-specific training would require separate permission, an appropriate data review and isolated evaluation.
What we will test
- Correct intent and policy selection.
- Answers supported by the supplied evidence.
- Handoffs for missing evidence, sensitive cases and requests outside authority.
- Separation between customers and protection of private service information.
- Useful response time and cost per successful task.
Voice and images,
when the workflow is ready.
Our longer-term direction includes customer-facing voice and multimodal support. Speech recognition and synthesis will be evaluated as separate components, with consent, retention terms, latency and human escalation tested before release. Image inputs require their own evaluation.
We aim for affordable operation through focused models and task-specific tools. Pricing and performance comparisons will follow measured results.
PLANNED CONTROL
Customer knowledge.
Customer permission.
The intended default is no retained conversation payloads and no training on customer conversations. Saved history, external tools, cloud processing and training will require explicit choices appropriate to the deployment.
Acadia will report uncertainty and hand off when needed. It will not be advertised as error-free or as a substitute for a customer’s service policies.
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.
Acadia 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. Acadia 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.