Deploy AI you can control.
Run capable models with your data, infrastructure and operating constraints in view.
Talk to BTLAI research + deploymentBuilt for ownership
Frontier AI. Efficient enough to own.
0103
Keep scrolling
One row of a real model. Every dot is a number it learned.
Store each number in
what it learnedwhat we keepwhat that costs5 possible values · average miss 0.0000
Hover any number
Reading the tensor sample…
- 22,020,096 values in the source tensor
- 14,000 values carried into the browser
- same quantization levels · no calibration corpus
Featured signal77.1 → 95.8behaviour retained
Representation
Same byte budget. Range selection changed what the bits could preserve.
Who it is for
Four groups ask different questions.
Build on models you can inspect.
Open-weight releases, native inference and systems that expose how the work gets done.
Explore the modelsBack the efficiency thesis.
Research that becomes models, systems and deployed capability.
Read the thesisMeasure what scale leaves out.
Measurements you can reproduce, the conditions they were taken under, and the runs that failed.
See the researchBTL Commercial
Contract the lab.
Governments and enterprises bring us problems that need more than an API. What comes back is research, a model built for their conditions, a system inside their infrastructure, or a capability that did not exist yet. Every contract names the bar it has to clear.
01 / Governments
Research and AI systems for institutions with sovereign constraints.
Models trained for local conditions, deployed on infrastructure you control, measured so you can see what they do.
Ministries · agencies · national programmesTalk to BTL02 / Enterprises
Applied AI research for organisations with hard technical requirements.
Built against your data and your constraints, with a baseline, a target and an acceptance test agreed before anything starts.
Banks · telcos · insurers · infrastructure operatorsTalk to BTLOne measured result
We did not add anything. We stopped wasting bits.
On a real weight tensor of 1.05 million values, the only thing we changed was how the available numerical range got allocated.
The model did not get larger. The byte budget did not change. No calibration dataset was used. The intervention took twelve GPU-seconds.
Behavioural retention moved from 77.1% to 95.8%. The companion instrument above lets you inspect the same failure directly in the tensor.
Read Range Before RepresentationFig. 01Aconditions attached
Behaviour retention77.1 → 95.8identical byte budget
Tensor1.05M values
Intervention12 GPU-seconds
Calibrationnone
Error reduction82.3%
The research map
We study capability efficiency.
It has seven forms. Each asks how much capacity a behaviour needs at a different point in the system. Two are measured. Five are not.
01RepresentationHow many bits are actually necessary to preserve behaviour?Measured
The first form we measured. Selecting the numerical range before choosing the representation moved behavioural retention from 77.1% to 95.8% at an identical byte budget.
artifact Range Before Representationcost twelve GPU-seconds
02ArchitectureHow much parameter capacity does a behaviour require?Measured
Small individually recoverable expert matrices regained more behaviour after compression than much larger dense matrices. The organisation of capacity changed how cheaply capability could be repaired.
artifact BRQresult 49.4% vs 23% recovery
03LearningHow much data and optimisation are required to acquire it?Open · next
This is where the cost of training is set. The question is whether behaviour-conditioned data selection can teach the same capability with fewer tokens and updates.
instrument not builtstatus dossier opened
04ModificationHow much of the system must change to alter one capability?Open
A capability that can be located should be repairable, replaceable, or addable without a full backward pass. The experiment has not been built yet.
instrument not builtstatus stage three design
05InferenceHow much compute must activate to execute it?Evidence only
BTL-4 Compact activates roughly 2.1B of 35.1B parameters per token. That is an architectural fact, not yet a measurement of the minimum compute a behaviour requires.
observation 2.1B active of 35.1Bstatus unstudied
06KnowledgeWhat belongs in weights, and what belongs in external memory?Dossier opened
Retrieval has reached a large parameter reduction in published work. The open question is where external memory becomes the more efficient representation for a real system.
adjacent system RetainDBstatus open
07DeploymentWhat hardware is actually necessary to serve it?Evidence only
A 35.1B model in a 9.96 GB artifact decoding at 31.9 tokens per second on a laptop says something about the floor. Nobody has measured where that floor is.
observation 31.9 tok/s on an M4status unstudied
Capability density is efficiency across all seven, measured in behaviour per unit spent.
The common question
How much behaviour does each unit of capacity buy?
We are trying to get more intelligence out of every bit, parameter, token and FLOP. A model spends bits to represent weights, parameters to store behaviour, tokens to learn, compute to reason, memory to retain information and hardware to run. We study that across seven forms and call the ratio capability density.
Capacity fieldselect a form
01 / RepresentationMeasuredHow many bits preserve the behaviour?
behaviour / bitrepresentation
behaviour / tokenlearning
behaviour / FLOPinference
behaviour / jouledeployment
behaviour / parameterarchitecture
behaviour / changemodification
Receipts
What we have actually measured.
The thesis is larger than the evidence. Two results exist today, with their conditions attached.
Fig. 02Representation
Range Before Representation
We did not add anything to the model. We stopped wasting the bits it already had.
- 77.1% → 95.8% retained
- same bytes
- no calibration corpus
Fig. 03Architecture
Behaviour Relearned Quantization
A 3M-parameter expert recovered more than twice as much behaviour as a 190M dense matrix. Structure mattered more than size.
- 3M experts → 49.4%
- 190M dense → 23%
- published ablation
The stack
Research should leave systems behind.
What survives the research becomes something we can build with.
Model family · ProposedExplore Tinfield
Tinfield
Open-weight models built around capability density.
The model family where the seven efficiency questions become design constraints from the beginning of training.
statusnot built
purposefrontier capability
shapeopen weights
The operating principle
Spend capacity where behaviour needs it.
The scaling era asks how much more capability we can obtain by spending more. BTL asks how much was necessary in the first place. If the measurements survive, intelligence gets denser: smaller, faster, easier to teach, easier to change and practical on far less hardware. Cost falls as a consequence. If they do not, we will know where the floor begins.