The Neural VM.
Ternary models run in chain state, in integers, byte for byte.
Add, subtract or skip
Three values per weight
In a BitNet-class model every weight is −1, 0 or +1. A matrix multiply stops being floating-point arithmetic and becomes additions, subtractions and skips over integer activations. Integers add the same way on every chip, in every order, so the result is exact.
- +1 adds the activation to the running sum
- −1 subtracts it
- 0 skips it: no work, no rounding
Integers don't drift
The same model on three unlike machines. In floating point, low bits differ by chip and batch; in pure integer, the hashes are one.
A pinned runtime
Every Miner runs the AIOS Module: pure-integer fixed-point kernels that give the same result whatever the batch size, with greedy or committed-seed decoding. It is pinned by content hash, so a node running anything else simply does not match. Cross-GPU floating-point byte identity is an open research problem; AIOS does not depend on it.
Small, and useful
What limits a model in chain state is state size and per-block compute, not model quality.
About 2 to 3 billion parameters
A natively trained ternary model of that size fits in under 1GB and is benchmark-competitive at its size. That is the credible ceiling for a model living in chain state. Larger models run off-chain across the Miner network and verify statistically.
- 01 / Load
Weights from chain state
The model is content-addressed state. Every node holds the same bytes.
- 02 / Execute
Integer kernels, pinned
The Neural VM runs the layers in the AIOS Module, deterministic by construction.
- 03 / Decode
Greedy or a committed seed
Sampling randomness is committed up front, so the output is reproducible.
- 04 / Commit
One output hash
Every conforming node lands on the same hash, and the committee can check it.
- Model class
- BitNet-class, 1-bit and ternary, integer-quantized
- Arithmetic
- Pure integer, batch-invariant kernels, no IEEE-754
- Decoding
- Greedy or committed seed, deterministic by construction
- Guarantee
- Byte-identical output on every conforming node
- Ceiling
- About 2 to 3 billion parameters natively ternary, under 1GB
- Never
- A frontier chatbot in consensus. Floating-point models verify statistically.
Same input, same bytes
That one property is what turns agreement into something a chain can check.