Expert-curated data · Disclosed architecture · Traceable claims
BASED
Truth, by design.
Based AI™ is a maximally truth-seeking intelligence, trained on curated data from leading experts in science, engineering, cryptography and economics. Every source is selected for signal. Low-signal web content and institutionally captured material are kept out.
fig. 02 — a dense body enters the field
Signal has weight.
- 01 / Models in the roster
- 07
- 02 / Architectures: MoE and Dense
- 02
- 03 / Models with published architecture and scale
- 100%
- 04 / OpenAI-compatible API
- 01
III — The thesis
An intelligence is only as reliableas the data it learns fromand the evidence it can show.
Based AI trains on curated, expert-authored corpora and records the provenance of every source. Answers arrive with their sources, uncertainty and reasoning visible, so the person using the model remains the final judge of what is true.
Models
Seven models.
One API.
fig. 04 — flower of life, nineteen circles of one radius
mercury●
MoE · 284B total / 13B active
Fast and efficient. Built for volume.
High-volume work at low cost: drafting, transformation, classification and the many small calls a production system makes every day.
venus
MoE · 1.6T total / 49B active
Deep reasoning for demanding work.
Longer reasoning chains and harder synthesis. When a task outgrows mercury, venus takes it on.
mars
MoE · 744B total / 40B active
General-purpose and reliable.
Balanced capability across analysis, code and writing. Steady performance under sustained load.
saturn
Dense · 27B
Long tasks and large documents.
Sustained generation across long inputs. Consistent output that stays on task from start to finish.
logos
MoE · 2.8T total / 104B active
Language and long-form reading.
Built for written work: long documents, close reading and prose that holds up to editing.
grug●
Dense · 27B
Short answers. Direct and plain.
Ask a question and receive the answer, stated plainly and briefly.
based●
MoE · 744B total / 40B active
The general default.
Our first choice for most work. Broad capability with a direct, sourced and measured style.
Platform
Data, models
and evaluation.
fig. 05 — the golden section, read from the outside in
- 01
Routing
One API. Seven models. Predictable behavior.
Every request goes to the same OpenAI-compatible endpoint and is served by the model you name. Each model publishes its architecture and scale, and the model you call is the model that answers.
- 02
Curated data
Expert-authored corpora, selected for signal.
We train on material written by leading practitioners in science, engineering, computer science, cryptography and economics. Sources are screened to exclude low-signal web content and institutionally captured material. Provenance is recorded at ingest, so every passage can be traced to its origin.
- 03
Training
The knowledge lives in the weights.
Each model is specialized on the curated expert corpus, so what it knows is learned during training and carried in its weights. The provenance of every source is recorded at ingest, and each answer states its claims, their sources and its uncertainty.
- 04
Finetuning
Your data. Your model.
We specialize a model on your material and serve it through the same API under its own name. It joins your roster with its architecture, data provenance and evaluation record documented.
- 05
Evaluation
Tested before release.
Every model and finetune is evaluated against task-specific harnesses: source-trace accuracy, claim support, calibration and factual accuracy. The harness is designed to find failure before users do.
Method
Ground Contact.
fig. 06 — Metatron's cube, thirteen centres and every line between them
01
Claim
Each output is stated as a claim that can be checked.
02
Source
Each claim links back to the material that supports it.
03
Test
Outputs are scored against task-specific harnesses before release.
04
Revise
Errors are identified quickly and corrected openly. Every answer remains open to revision.
Evaluation
Tested before release.
fig. 07 — the sieve: every composite is struck by an arc, what survives is prime
- 02
Source-trace accuracy
Does each answer cite the material it came from?
- 03
Claim-support rate
What share of claims are supported by their cited sources?
- 05
Calibration error
Does stated confidence match observed accuracy?
- 07
Refusal quality
Does the model decline the right requests for the right reasons?
- 11
Factual accuracy
Are stated facts correct against the expert record?
- 13
Reasoning consistency
Does the reasoning hold when the question is asked again?
fig. 08 — what cannot be seen whole can still be measured, one projection at a time
Principles
Five rules we build by.
fig. 09 — Apollonian gasket, circles tangent to circles without end
- 01
Disclose the model
Every model publishes its architecture, parameter scale and intended use. A system has to be described before it can be evaluated.
- 02
Curate for signal
Training data comes from expert-authored sources and is screened for low-signal and institutionally captured content.
- 03
Show the evidence
Every claim carries a path to its source material. Checking an answer should be easier than accepting it.
- 04
Correct errors openly
Our models do not claim certainty. We build systems that surface errors early and record each correction.
- 05
The user decides
Models reason, summarize, analyze and critique. Final judgment stays with the person using them.
V — Access
Build with Based AI.
API access, custom builds and finetunes start with a short request. Tell us what you are building and a member of our team will reply.