BitBrain Model Platform

AI that works
where others can't.

Modern AI runs on floating point — billions of approximate multiplications. We train neural networks whose weights are small whole numbers, so inference is integer addition and bit-shifts. The models are radically smaller, run on almost anything — and because whole-number arithmetic is exact, a result can be replayed bit for bit, anywhere.

Zero multiplications
25 KB smallest production model
Bit-exact replayable inference
See what we're building

What we build.

Each BitBrain model is trained from scratch with our proprietary pipeline, on whole-number weight alphabets tuned to the task. We train the model. You license and deploy it.

Demonstrated

Vision & Image

Lightweight whole-number vision models for object detection, recognition, and automated inspection — accurate without a float multiplier in sight, small enough for continuous camera processing on the device itself.

Status Demonstrated across real domains: road-damage detection from a vehicle camera (a 398 KB model at 24 ms/frame on CPU), aerial crop surveying, face verification, and medical-imaging research on a published clinical benchmark.
Mobile cameras IoT sensors Quality control Agriculture
Demonstrated

Language & Sequence

Language and time-series models in two complementary modes: full-context models for writing and understanding, and streaming models whose cost per step is constant. Both run on the same whole-number arithmetic, and both are small enough to run client-side.

Status A 30M-parameter model writes fluent English entirely in adds and shifts — billions of integer operations per paragraph, zero multiplications — with a cryptographic replay receipt on every generation. Larger models are in training.
Predictive input Anomaly detection Telemetry
Custom

Your Problem

Our training pipeline is model-agnostic. If a neural network can solve your problem, we can build a BitBrain version of it — dramatically smaller, deployable on your target hardware.

01 Define the problem & constraints
02 We train a BitBrain model
03 You license & deploy
Discuss your project

Measured, not promised.

Public benchmarks where they exist, published human baselines where they exist — and the caveats stay attached. Live demonstrations of every model family are available on request.

95.48%

Keyword spotting, 25 KB

Google Speech Commands v2, official test set — above the published floating-point reference model (94.4%), in a model that fits in a microcontroller's spare memory.

2.6 PESQ

Noise suppression, 88.5 KB

Standard VoiceBank-DEMAND speech-quality benchmark, protocol-strict. Best in class at this size and power budget — large float research systems score higher and we don't claim otherwise; they don't fit where ours runs.

24 ms

Road damage, per frame on CPU

A 398 KB detector finding road damage from a moving vehicle camera — no GPU anywhere in the pipeline. Proof of concept.

8 vs 1

Melanoma reader study

On a published clinical benchmark's reader study, our sealed-test model scored above the average of the eight dermatologists on the same 100 images. A narrow, research-only claim — not a medical device — but the evaluation discipline behind it is the product.

20 / 20

Analog circuit agreement

A complete model reproduced in SPICE circuit simulation, 20/20 test cases in agreement — and simulated analog memory hardware drifted just −0.16% over a year. The arithmetic is already at home in physics.

0

Multiplications

Our latest language models run the entire stack — attention and normalisation included — without a single multiplication, counted live by the browser's own op counter: billions of adds and shifts per paragraph, zero multiplies.

Why it works.

Extreme quantisation should destroy model accuracy. Ours doesn't. A proprietary training pipeline makes the difference.

01

Quantised from birth

BitBrain models aren't compressed after training — they're trained quantised from the first step, on weight alphabets of small whole numbers: as few as three states {-1, 0, +1}, at most five {-2 … +2}. Every forward pass is integer addition and bit-shifts; our latest language models run the entire stack without a single multiplication.

02

Novel error correction

Quantisation introduces systematic errors that compound through layers. We developed a proprietary correction mechanism that operates during training — continuously detecting and compensating for quantisation noise without adding inference cost. The result: accuracy that standard quantisation can't match.

03

One platform, two modes

Full-context models give language tasks deep understanding of everything in the window. Streaming models process unbounded live signals — audio frames, camera feeds, telemetry — at constant cost per step, with state measured in kilobytes. Same whole-number arithmetic, same export pipeline; we pick the mode per task, and some products use both.

04

Deploy anywhere

Models export to a single portable file. Inference is pure integer arithmetic — runs in any browser, on any microcontroller, in any language. No framework dependencies. No GPU. No runtime. The same model runs on a cloud server and a hearing aid.

Every run carries a receipt.

Whole-number addition is exact, so in deployment form a BitBrain model produces the same bits on any hardware, any thread count, any platform — and every output can carry a cryptographic replay receipt. Re-run it anywhere; the hashes match. In regulated and safety-critical settings, that is the difference between "trust us" and "check for yourself."

BitBrain, deployed form

The same model verified bit-for-bit across three independent implementations — a GPU server, a laptop, and a web browser.

  • Golden receipts reproduce hash-for-hash, 4/4
  • 0/40 replays diverged under shuffled arithmetic order
  • Streamed and batched inference: identical bits
  • An audit trail a regulator can re-run

An identical float twin

The same architecture built on conventional floating point, put through the same test.

  • 32/32 receipts broke
  • Reordering the additions changed the output
  • Roughly 1 in 3 replays produced different results
  • None of its own golden runs reproduced

Not post-training compression.

Conventional approach

Train a large float32 model, then compress it afterwards. Quality degrades. The model wasn't designed for the constraints.

  • Train at full precision
  • Quantise after training
  • Accept quality loss
  • Still needs GPU for training

BitBrain approach

Train directly in the target precision from step one. The model learns to be accurate within its constraints.

  • Quantised from first gradient
  • Error correction throughout
  • No post-hoc degradation
  • Designed for the target hardware

Where this goes.

Our technology isn't just about smaller models today. The same properties that enable extreme quantisation unlock entirely new deployment paradigms.

Now

Digital deployment

BitBrain models running on standard hardware today — browsers, microcontrollers, edge devices — with measured public-benchmark results on audio, vision, and language, and replay receipts built in.

Production ready Licensable
Near term

Scale & serve

Streaming models keep per-session state in kilobytes, so modest CPU servers — or the user's own device — replace GPU fleets for live workloads. Larger language models are in training on the same recipe, with the audit trail built in.

CPU-only serving On-device
Horizon

Analog silicon

A multiply-free, noise-resilient architecture maps directly onto analog circuits: a whole-number weight is a wire, a sum is currents meeting at a node. Physics does the compute — already validated in circuit simulation, with simulated analog memory drifting under a fifth of a percent per year.

SPICE validated Analog inference Zero FPU

Interested in efficient AI?

Whether you need a model for a specific problem, want to license an existing BitBrain model, or are exploring what efficient AI could do for your product — we'd like to hear from you. Live demonstrations of every model family are available on request.

Custom model training and per-deployment licensing available.