krummholzThe weights come
with field notes.
A general-purpose model in beta: Atlantic v8 (27B) for desktop or hosted use, and Atlantic Lite v8.1 (4B) for phones and low-end hardware. Both are built to be truth-seeking first: they say "I don't know" rather than guess. Permissioned tools and user-reviewed dynamic weights remain part of the product plan.
Public beta files for testing and feedback. File sizes are not RAM requirements. Atlantic Lite Q4_K_M has been scored separately; Atlantic Q4_K_M evaluation and phone acceptance tests remain pending. Runtime guards are separate from the bare weights.
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.
Atlantic v8 and Atlantic Lite v8.1 are released with open weights under Apache 2.0. Dynamic, user-reviewed learning is in development. Hosted inference and training on your private data will be paid services. Third-party components retain their applicable licenses and notices. Beta weights are for testing and feedback; the model cards list files, results and limits.
MODEL BETA RELEASES
Two sizes.
One family.
Atlantic v8 (27B)
Fine-tuned from Qwen3.8-27B for workstations, desktops and hosted use. Released November 15, 2026. The GGUF weight files are about 17 GB (Q4_K_M) or about 19 GB (Q5_K_M), with a separate vision projector and the LoRA adapter.
Atlantic Lite v8.1 (4B)
Fine-tuned from Qwen3.5-4B for phones and low-end hardware. Released November 15, 2026. The GGUF weight files are 2.7 GB at Q4_K_M (for phones) and 3.1 GB at Q5_K_M, plus a full-precision bf16 file, a separate vision projector and the adapter.
The listed sizes describe files, not required RAM. Atlantic Lite v8, released November 14, 2026, stays available on Hugging Face. Versions 6 and 7 are withdrawn because the newer versions are more accurate and safer. Acadia and Katahdin will be distilled from Atlantic v8; our earlier 35B mixture-of-experts core is deprecated.
Atlantic v8 model card · Atlantic Lite v8.1 model card · How version 8 was trained, in detail →
Measured results
Fewer wrong answers.
More honest ones.
These models are not built to be maximally helpful. They are built to be truth-seeking 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 assist with harm.
We ran the same evaluation on the old and new versions, with full-precision weights and 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, never used for training).
Atlantic 27B: v7 to v8
- Fact questions without tools: wrong answers fell from 49% to 16%. Right answers went from 50% to 48%, and "I don't know" from 1% to 35%.
- With tools: it searches on 81% of fact questions, up from 51%, and gave no wrong answer among the 44 it answered without searching.
- Reasoning that never finished: 34% to 0.
- Tool calls well formed: 60 of 60.
- Long prompts where it repeats itself or runs into the limit: 25 to 0 of 120.
- HarmBench harmful replies, model alone: 1 to 0 of 300.
- Corrects a wrong calculation stated as fact: 59 of 60. Corrects a wrong fact offered for confirmation: 10 of 30.
- Probes passed: 172 to 168 of 199.
- Q4_K_M file: being measured.
Atlantic Lite 4B: v8 to v8.1
- Fact questions without tools: wrong answers went from 3% to 2% (version 7 was wrong on 52%). Right answers rose from 28% to 33%, and "I don't know" from 49% to 62%.
- With tools: it searches on 91% of fact questions and gave no wrong answer among those it answered without searching.
- Reasoning that never finished: 20% to 3%.
- Tool calls well formed: 60 of 60.
- HarmBench harmful replies: 24 to 10 of 300 for the model alone, and 2 to 0 of 300 behind the Krummholz harness. The harness rules were written on half of HarmBench, so that figure is partly in-sample.
- Corrects a wrong calculation stated as fact: 29 to 39 of 60. Corrects a wrong fact offered for confirmation: 7 to 10 of 30.
- Harmless requests refused: 4 to 19 of 120, a trade we accepted. Probes passed: 132 to 117 of 199.
- 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.
Known weak spots: pushback on false facts is still low, probe scores dipped, and more harmless requests are refused. We treat refusing a safe request as the lesser evil against accepting a harmful one. Atlantic Lite must be run behind the harness or an equivalent filter; the harness source is not published yet. The test set is small, so read these numbers as early evidence.
Your task.
A clear boundary.
Atlantic 27B runs on capable desktops or hosted servers. Atlantic Lite, at 4 billion parameters, is our compact model for phones and low-end hardware; our phone acceptance tests have not yet run on it. Task routing, offline scope and the initial tool set remain under evaluation; hosted processing will be clearly disclosed.
Tool use can delegate bounded work to purpose-built functions, but it does not automatically recover capability lost through model compression. We will test task quality, local speed, memory, context size, hosted latency and cost before setting supported hardware and release limits.
Understand a document before acting.
Ask for a source-backed summary of a document on your device. Atlantic would point to the relevant material and flag uncertainty. A request to take an external action would pass through separate permission checks and any required approval.
This is an intended pilot workflow, not a demonstration of an available app.
The model proposes.
The application checks.
The planned tool layer checks identity, permissions, arguments and usage limits before execution. We will start with read-only capabilities and individually authorized writes, with scoped access and a stop control.
Text from a website, email or document is information to analyze, not authority to change a user’s instructions. Adversarial and prompt-injection tests will be part of the release review.
Unrestricted computer control and broad autonomous campaigns are outside the initial pilot.
What we will test
- Completion of defined tasks and correct tool selection.
- Refusal or escalation when authority or evidence is missing.
- Resistance to adversarial instructions in retrieved content.
- No unauthorized execution or cross-user disclosure in the defined release test suite.
- Local and hosted behavior, offline limits, memory, latency, heat and battery use on the target phone.
Broad usefulness.
Documented foundations.
Version 8 continues from the version 7 weights with rounds of preference training. In each round the model answers questions with known answers, code labels its replies, and the model learns to prefer its own better reply over its own worse one. Apart from a few fixed opening words, no training text in version 8 was written by a person or by another model.
The questions and graders come from Wikidata (CC0), Natural Questions open (CC BY-SA 3.0) and GSM8K (MIT). One disclosure carries over: version 7 included math and code solutions written by DeepSeek-R1 and QwQ-32B, outside our policy of training only on text from American open-weight models. Version 8 adds no such text, but that material is still in its lineage. Read the full training notes.
Publish the limits
alongside the capability.
The model cards publish lineage, files, evaluation results and limits for each release, along with its license and applicable upstream terms. Supported hardware will follow once phone acceptance tests have run.
Security and accuracy are goals to test, not absolute guarantees. Modifiable offline models cannot enforce central usage quotas or immutable safety behavior; hosted service limits will be described separately.
MODEL PERSONALIZATION / IN DEVELOPMENT
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 →BETA RELEASES
The early launch release is out.
Full launch: January 2, 2027.
Atlantic v8 and Atlantic Lite v8.1 were released with open weights on November 15, 2026 (Atlantic, Atlantic Lite). Both are still beta. The full launch is planned for January 2, 2027.
What is still open
- Under 1% hallucination: not reached. Without tools, Atlantic 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.
- Harmful requests: 0 of 300 for Atlantic v8 alone and for Atlantic Lite v8.1 behind the harness; Atlantic Lite alone, 10 of 300. The harness source is not published yet.
- Consumer hardware: the Atlantic Lite Q4_K_M file is 2.7 GB, but phone acceptance tests have not run yet.
- Dynamic weights personalization: in development, not released.
- Paid hosted inference and private-data training: planned.