krummholzAI
Beta · v8 and v8.1
Open-weight models for customer service, forestry, fisheries and general-purpose work. Built to give individuals and organizations control over their AI, not lock them into Big Tech.
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.
Two general-purpose models you can download and run today, and two specialists in development. Both released models have open weights under Apache 2.0. Versions 6 and 7 are withdrawn because the newer versions are more accurate and safer.
Krummholz models are not built to be maximally helpful. They are built to seek the truth and to hold to a written constitution: a collaborator, not an assistant. They say "I don't know" rather than guess, and decline rather than help with harm. The constitution is printed in full on the Atlantic v8 and Atlantic Lite v8.1 model cards.
27B · workstations, desktops and hosted use
Questions, writing, planning and document work, with room to reason. Ships with a vision projector for images.
4B · phones and low-end hardware
The same Krummholz training in a model small enough for a phone. Ships with a vision projector for images. Run it behind the Krummholz harness or an equivalent filter.
Atlantic Lite v8 (November 14, 2026) stays available: Lite v8 on Hugging Face.
Support that knows when to hand off
Answers grounded in a company’s approved policies, request classification and clear escalation to people. The first pilot covers one text-based service workflow. Still in development.
Domain knowledge, working-world context
A planned assistant for understanding field records, finding source material and explaining reviewed analyses, paired with validated analytical tools. Still in development.
Benchmarks
Each row shows the new release, with the previous version underneath. Old and new took the same evaluation, on full-precision weights with reasoning on: 229 held-out fact questions asked without tools and again with tools, 199 probe questions, 120 long prompts, 60 tool-use turns, 120 false claims and HarmBench (300 harmful behaviors, official classifier, never used for training).
27B · v8, compared with v7
Q4_K_M file: being measured.
4B · v8.1, compared with v8
Q4_K_M file (phones): without tools 24 wrong (10%), 88 right and 108 "I don't know" of 229; with tools no wrong answer; HarmBench 6 of 300 for the model alone and 0 of 300 behind the harness.
For scale, Atlantic Lite v7 was wrong on 52% of fact questions without tools. With tools, Lite v8.1 searches on 91% of fact questions and Atlantic v8 on 81%; neither gave a wrong answer among those it answered without searching (44 answers for Atlantic v8). The harness rules were written on half of HarmBench, so the harness figure is partly in-sample. Lite refusing more harmless requests is a trade we accepted: refusing a safe request is the lesser evil against accepting a harmful one. The model cards have the full results: Atlantic v8 · Atlantic Lite v8.1.
Release goals
Without tools, Lite v8.1 is wrong on 2% of fact questions and Atlantic v8 on 16%. With search tools, neither gave a wrong unsearched answer in this test. The test set is small (229 questions).
60 of 60 well-formed tool calls on both models.
0 of 300 for Atlantic v8 on its own and for Lite v8.1 behind the harness. Lite on its own: 10 of 300, so Lite must run behind the harness or an equivalent filter. The harness source is not published yet.
The Lite Q4_K_M file is 2.7 GB. Phone acceptance tests have not run yet.
User-reviewed learning is in development and not released.
Pushback on false facts (10 of 30), probe scores dipped, and more harmless requests are refused.
After reading these answers we fixed five gaps in our scorer and runtime, and both model cards say so. The model cards list every suite and result. Read the model cards →
MODEL PERSONALIZATION / PLANNED PERSONALIZATION
Dynamic weights describes our planned user-reviewed learning: models adapt only through approved changes. The mechanism and architecture remain under evaluation.
“What I learned today” will let users accept, reject or edit proposed learning. Retention and training require separate opt-in choices. Business administrators set who can approve changes.
Approved learning may make your local model differ from hosted versions. Version information will explain those differences. Local learning will not automatically be shared with hosted models.
We aim to carry approved learning into future model versions. Compatibility, backup, migration and rollback still need testing.
Planned safeguards include source tracking, untrusted-input isolation, bounded updates, adversarial tests and rollback. They reduce risk, not eliminate it. Learned preferences never grant tool permissions.
In development: Learning review and update preservation are not released and may follow the first pilot.
Company update: Training continues for now. Read Hayven’s note about Belfast and what comes next.
Version 8 starts from the v7 weights and adds rounds of preference training (DPO) with reasoning on. In each round the model answers questions with known answers, and code (not a person, not another model) labels each of its own replies: right, wrong, a clean "I don't know", a well-formed tool call, a correction of a false claim, and so on. The model is then trained to prefer its own better reply over its own worse one. Apart from a few fixed opening words, no training text in v8 was written by a person or by another model.
New in this release: short, plain refusals of harmful requests, pushback on false claims built by code, finishing with "I don't know" instead of running out of reasoning, and not calling a tool that was not offered. A control run with shuffled labels did not improve, so the gain comes from the labels.
Disclosure carried over from v7: math and code solutions from public NVIDIA datasets, written by DeepSeek-R1 and QwQ-32B, made up about 13% of Atlantic Lite v7’s training examples and about 29% of Atlantic v7’s smaller set. That is outside our policy of training only on text from American open-weight models. Version 8 adds no such text, but it continues from the v7 weights, so that material is still in its lineage.
The long version of how v8 was trained, for fellow nerds →
Atlantic v8 model card → · Atlantic Lite v8.1 model card → · All weights on Hugging Face →
Our business model pairs open weights with paid hosted inference and private-data training for companies and individuals. Hardware support, offline scope, routing and hosted limits will follow testing. Private-data training requires explicit permission and separate terms.
Inspect our preliminary text-data pilot →BESPOKE AI TRAINING
Discuss a scoped training project using data you are authorized to use. We start with the task and evaluation criteria, compare fine-tuning with retrieval and other options, and plan around your deployment requirements.
Explore bespoke training →We plan transient processing by default. Saved history, retrieval indexes, cloud processing and private-data training will have separate, explicit choices.
Retrieval can supply current policies without retraining. Training on private data is a separate, opt-in service.
Our synthetic datasets support model development and customer experiments, with generated content clearly labeled.
We favor source-backed answers, visible uncertainty and escalation when evidence is insufficient. Evaluations will report known limits.
Documents and websites are untrusted inputs. Permission checks stay outside the model. We will disclose hosted processing and identify what works offline.
Our neutrality goal is consistent evidence standards and fair treatment of relevant perspectives, assessed through defined evaluation cases.
The early launch release is out: Atlantic v8 and Atlantic Lite v8.1, open weights. Full launch: January 2, 2027. The models are still in beta. Further releases depend on evaluation, privacy, security and license review.
Our LLMs are released with open weights under Apache 2.0. Third-party base models and components retain their applicable licenses and notices. Code and datasets keep their own licenses.
Current status: Atlantic v8 and Atlantic Lite v8.1 are in beta, released November 15, 2026. Versions 6 and 7 are withdrawn. The 35B core is deprecated; Acadia and Katahdin will be distilled from Atlantic v8, are not trained yet and have no release date. No hosted service is available yet. Dynamic weights are in development and not released.