← Products
Model 02 — In Research

GYRE

Coming soon.

A looped, adaptively-halting MoE-SSM with exact retrieval and contracted compute budgets. 3.03 billion parameters, built by architectural surgery — an attempt at the fast-and-exact corner of the design space that attention cannot reach.

The architecture is measured and verified. The trained model is what comes next.

100%
exact retrieval
at 12M tokens
5.24M
tokens / forward pass
one 96 GB GPU
75.5×
faster than attention
at 1M tokens, mixing path
1.000
MQAR, zero training
transformer scores 0.000

Why it could change the economics

Four structural
differences

01

No attention. No KV cache.

Sequence mixing is a linear-time state-space recurrence. Exact long-range recall comes from a separate retrieval organ — not softmax over every pair of tokens.

02

Depth bought with iteration.

Four core blocks are looped 1–8 times under a learned halting policy, reaching an effective depth of 36 layers from 8 unique blocks. Test-time compute at zero extra parameters.

03

Active parameters are a contract.

A prompt selects one of 16 declared tiers and the model activates exactly that many parameters — 0.673B to 2.482B out of the same 3.03B resident weights. Declared, auditable, provably hit.

04

Retrieval is exact.

A streaming suffix automaton resolves verbatim context matches in O(m) time independent of context length, and can feed the output distribution directly.

The honest limit

GYRE has never been trained to convergence. It has no scores on general benchmarks yet. Everything shown describes an architecture that is measured, verified and ready — not a model that has been proven. Whether it changes the industry is decided by a training run that has not happened.

That is exactly why it says Coming Soon.

Want to know when GYRE ships?