CA3 Contact
CA3

Your model learns every day.
You decide what it keeps.

CA3 lets your open model keep learning after deployment, by your rules. What it learns is kept alongside the original, which never changes.

Models

AI models tested with CA3,
and their applications.

Every use case, grouped by type of model: from 22 million to 122 billion parameters, by size and by who makes it, all of them open-weight. The hollow dot ran without CA3.

Technical testsHome experimentsSame pilot, open model alone, without CA3
Every use case of CA3 and its models, grouped by type of model, by size and maker; the hollow dot, the first phase of the Biome Makers pilot, ran without CA3 Dot plot, one row per use case, grouped by type of model, model size in parameters on a log scale from 10 million to 300 billion; under the axis, the maker of each model. The brain-implant decoder and the recommender have no published model size. 10M 100M 1B 10B 100B Size, parameters Made by Meta NASA · IBM Google Multiverse Alibaba Sony AI · UIUC Alibaba Hugging Face IBM AWS Alibaba Ai2 Alibaba Meta Alibaba Krea Alibaba NVIDIA OpenAI Alibaba LANGUAGE MODELS · LLMs SPEECH AND AUDIO MODELS VISION AND IMAGE MODELS EARTH-OBSERVATION FOUNDATION MODEL ROBOT FOUNDATION MODEL · VISION-LANGUAGE-ACTION NEURAL DECODER · BRAIN IMPLANT RECOMMENDER AI cybersecurity Biome Makers · client memory Biome Makers · farm reports Self-healing Language AI judges Model edits Flip-flop Home robot Personal app Home library Speech Sound Vision Image generation Short film Satellite Robot arm Brain implant Recommendation SPINT decoder · size unknown, not given in its paper Matrix factorisation · our own small model, under 1 million parameters gpt-oss-120b: 116.8 B parametersgpt-oss-120b Qwen3.5 122B-A10B: 122 B parametersQwen3.5 122B-A10B Qwen3.8 27B: 27 B parameters, open model alone, without CA3Qwen3.8 27B · no CA3 Qwen3.5 0.8B: 800 M parametersQwen3.5 0.8B Qwen2.5 0.5B: 500 M parametersQwen2.5 0.5B Qwen2.5 1.5B: 1.5 B parametersQwen2.5 1.5B · SmolLM2 1.7B SmolLM2 1.7B: 1.7 B parameters Qwen2.5 7B: 7 B parametersQwen2.5 7B · Llama 3.1 8B · Qwen3 8B Llama 3.1 8B: 8 B parameters Qwen3 8B: 8 B parameters Qwen3 14B: 14 B parametersQwen3 14B Nemotron 3 Nano 30B: 30 B parametersNemotron 3 Nano 30B Strands Decider 2B: 2 B parametersStrands Decider 2B Qwen3.5 9B: 9 B parametersQwen3.5 9B Gemma 3 270M: 270 M parametersGemma 3 270M Qwen3.5 0.8B: 800 M parametersQwen3.5 0.8B Qwen3.5 0.8B: 800 M parametersQwen3.5 0.8B Qwen3 1.7B: 1.7 B parametersQwen3 1.7B LittleLamb 0.3B: 300 M parametersLittleLamb 0.3B Qwen3 4B: 4 B parametersQwen3 4B Qwen3 4B: 4 B parametersQwen3 4B Granite Speech 4.1 2B: 2 B parametersGranite Speech 4.1 2B MMAudio L 44k: 1.03 B parametersMMAudio L 44k DINOv2-small: 22.1 M parametersDINOv2-small Krea 2: 12 B parametersKrea 2 Krea 2: 12 B parametersKrea 2 Prithvi-100M: 100 M parametersPrithvi-100M MolmoAct2: 5 B parametersMolmoAct2
10M100M1B10B100B

Language models · LLMs

AI cybersecurity

gpt-oss-120b (OpenAI)

Biome Makers · client memory

Qwen3.5 122B-A10B (Alibaba)

Biome Makers · farm reports

Qwen3.8 27B (Alibaba), without CA3

Self-healing

Qwen3.5 0.8B (Alibaba)

Language

Qwen2.5 0.5B (Alibaba) · Qwen2.5 1.5B (Alibaba) · SmolLM2 1.7B (Hugging Face) · Qwen2.5 7B (Alibaba) · Llama 3.1 8B (Meta) · Qwen3 8B (Alibaba) · Qwen3 14B (Alibaba) · Nemotron 3 Nano 30B (NVIDIA)

AI judges

Strands Decider 2B (AWS) · Qwen3.5 9B (Alibaba)

Model edits

Gemma 3 270M (Google) · Qwen3.5 0.8B (Alibaba)

Flip-flop

Qwen3.5 0.8B (Alibaba)

Home robot

Qwen3 1.7B (Alibaba)

Personal app

LittleLamb 0.3B (Multiverse) · Qwen3 4B (Alibaba)

Home library

Qwen3 4B (Alibaba)

Speech and audio models

Speech

Granite Speech 4.1 2B (IBM)

Sound

MMAudio L 44k (1.03B, Sony AI · UIUC)

Vision and image models

Vision

DINOv2-small (22.1M, Meta)

Image generation

Krea 2 (12B, Krea)

Short film

Krea 2 (12B, Krea)

Earth-observation foundation model

Satellite

Prithvi-100M (NASA · IBM)

Robot foundation model · vision-language-action

Robot arm

MolmoAct2 (5B, Ai2)

Neural decoder · brain implant

Brain implant

SPINT decoder, size unknown (not given in its paper)

Recommender

Recommendation

Matrix factorisation, our own small model, under 1 million parameters

10M100M1B10B100B

Model size, parameters, log scale

Sizes as published by each model's maker; point at a dot for its name and size. The hollow dot is the first phase of the Biome Makers pilot: the open model wrote the farm reports on its own, without CA3. The brain-implant decoder (SPINT) has no size in its paper, and the recommender is our own matrix factorisation, under 1 million parameters, below the axis: their rows carry no dot.

Open23models

open-weight, from 22 million to 122 billion parameters. On your servers or ours.

See the models →
Safe0of 24

planted falsehoods kept. A RAG over the same news repeated 8.

See the case →
Fast8×

faster answers than the same model without CA3: 0.51 s against 4.03 s, in a personal app.

See the case →
Accurate+43points

on six new European regions in six days, for NASA and IBM's crop model.

See the case →

Every model stops learning
the day it ships.

Retrain it, and it can break what already worked. Add RAG, and it pays to reread the same context on every question. Either way, nobody sees what changed.

Learns every day.

Keeps what it knew.

Blocks bad data.

Forgets on request.

You decide.

Technical tests

The same control.
Fifteen technical tests.

From agents that pass each other notes, AI judges that learn and a 122-billion-parameter model on a company's client data to a brain-implant decoder: language, robots, satellites, images, speech, vision, sound and recommendation.

The 100 answers the edit broke, turning right as the model heals: 98 of 100
Self-healing · new article

Sparks of Self-Healing

The model heals itself; CA3 makes sure nothing else breaks. An AI edit made a small model six times better at math and quietly broke a skill. With CA3 as its controller, the model found the damage and healed it: 76.3% → 98.5%, with the math gain untouched.

Alibaba Qwen3.5-0.8B · edit by GPT-6 Astra, MIT CSAILRead the article →
120 runs per defense, contagions in red: no defense 40.0%, warning 37.5%, note auditor 29.2%, CA3 10.0%
AI cybersecurity · agents · new

75% less risk in attacks like OpenAI's.

Agents passing notes, like the ones that attacked Hugging Face. A rigged note fools the next agent 4 times in 10; with today's defenses, 3 or 4; with CA3 checking every note, 1.

OpenAI gpt-oss-120b · open weights · one GPUSee the case →
The assistant's panel, redrawn, with two answers quoted from the run
Agronomy · Biome Makers pilot · new

Answers from each client's own data.

Real data of two clients. CA3 keeps one memory per client and region, learns new sheets on top and removes one on request. Account questions go to a program that computes every figure: none wrong in 130 answers, against 72 for document search.

Alibaba Qwen3.5-122B-A10B · open weights · two GPUsSee the case →
Language

Learns the news every day.

More right than RAG, at a seventh of the compute. Not one planted falsehood kept.

Alibaba Qwen · Meta Llama · SmolLM2 · NVIDIA Nemotron · 0.5B to 30BSee the case →
Of every 100 checks: 66 settled by the small judge, 27 by the big judge, 7 by a person
AI judges · new

A person reviews only 7 checks in 100.

Every new fact a model learns is checked. A small judge learns from the big one to settle two thirds of the checks, and loses none of the 42 planted falsehoods the big one stopped.

Qwen3.5-9B and Strands Decider 2B · open weights · one GPUSee the case →
Brain implant

Learns each new day.

Seven new days learned; the earlier days kept at 0.939, fine-tuning 0.712.

SPINT decoder · FALCON public recordingsSee the case →
Robot arm

Learns new tasks, keeps its job.

All ten everyday tasks kept; ordinary fine-tuning lost them all.

Ai2 MolmoAct2 · simulatedSee the case →
Satellite

Learns Europe, keeps the US.

+43.2 points in six regions; the US held at 60.

NASA and IBM Prithvi-100MSee the case →
Image generation

Learns a brand's characters.

One new character a day; none forgotten, none mixed up.

Krea 2 · research demonstrationSee the case →
Model edits

Uncovers hidden problems before deployment.

7 weight edits written by an AI: 4 held back, each with what it breaks. 3 of those 4 had passed the usual benchmarks.

Alibaba Qwen3.5 · Google Gemma 3 · edits written by an AISee the case →
Model edits

Keeps both versions.

The original and an AI edit in one model, flipped in an instant; back to the original, bit for bit.

Alibaba Qwen3.5-0.8B · CA3 Flip-FlopSee the case →
Speech

Learns a team's vocabulary.

Names 8% → 97%, and its English kept.

IBM Granite Speech 4.1See the case →
Vision

Learns a new camera.

41% → 87% on it, the old camera kept, a planted trigger refused.

Meta DINOv2See the case →
Sound

Learns each client's sounds.

39.8 MB per client, on one shared model.

Sony MMAudio · research demonstrationSee the case →
Recommendation

Learns new ratings.

Every fake account's ratings held back.

Matrix factorisation · MovieLens-1MSee the case →
Your model

Test it on your model.

Four to eight weeks on your own data, under NDA, with the same report.

Contact →
Home experiments · CA3 Labs

The same control,
at home.

Everyday things that learn a family: a robot vacuum, a coach app, a bookcase and a film with the toys of the house.

Control panel

It asks you only when in doubt.

The routine it settles on its own. The doubtful cases come to you, with the evidence side by side, and every decision stays on record.

The CA3 control panel on day 40 of the recorded test
93 in 100checks settled on its own
7 in 100sent to a person

229 of 3,136 checks, on three language models · every day a version: 1 of 29 undone on its own · See the case

Efficiency

Learn it once.
Pay less on every answer.

7×less compute per answer than RAG, answered from what it learned, with no lookup.
$2–4of GPU a month for one customer's daily learning, in the recorded test.
40%of the learning time to carry what it learned to a new base model, instead of starting over.

From the language test: Alibaba Qwen3 14B, round 20d; moving to a new base model, Qwen2.5 7B to Qwen3 8B Base, round 13 · See the case

Pilot

Test it on your model.

Four to eight weeks on your own data, under NDA. You get the same report on your data: accuracy and cost against your RAG, what it learned, what it refused and what it forgot.

We use these details only to reply to you.

Choose your modelOne from the tested list, or yours from the same families.
Connect your dataDocuments, questions and corrections, as they arrive.
Set your rulesWhat it learns on its own, and what goes to a person.
Run itOn our servers or yours. Your data and what it learns stay yours.
Read the reportAgainst your RAG, with every decision on record.