AT-1 Mixture-of-Specialists

Every top open model, one verified vault — the right specialist per query.

A Mixture-of-Experts model only fires a few experts per query — the rest sit idle, taking up RAM. AT-1 packs every specialist into one compressed, SHA-256-verified container, routes each query with a signed router, and streams only the routed expert. The result is byte-identical to the full model, with a working set of one expert instead of all of them.

One verified vault

Every specialist's weights live in a single content-addressed, compressed container. Each expert's tensors are individually addressable and SHA-256-hashed; the whole container has a root hash.

Signed router

The routing prototypes are Ed25519-signed, so a tampered router or expert set fails verification. Every routing decision emits a signed receipt binding the query to the expert that answered it — provable, auditable routing.

Stream one expert, not N

Only the routed expert's tensors are decompressed and verified per query, hot-cached for reuse. A model that would need all experts resident now runs with a working set of one — the streamed weights are byte-identical to the full model.

Editable, still verified

Add or remove a specialist and the container is rebuilt and re-signed. verify() re-checks every expert's hash, the container root, and the router signature — so the vault is always provably intact.

Pack, route, stream — one command to see it

The demo packs specialists, routes a query to the nearest one, streams just that expert, proves the streamed weights are byte-identical to the source, and verifies the whole container.

at1 moe demo
#  packed 4 experts, routed query -> expert 2 (score 0.99)
#  streamed expert byte-identical to source: True
#  working set after one expert: 16,384 B (of 4 experts)
#  verify: PASS  (8 tensors, sig OK)
#  after add/remove: experts=4, verify=PASS

Always provably intact

verify re-hashes every expert tensor, re-checks the container root, and verifies the router signature. Tamper with an expert and its hash fails; move a routing centroid and the signature fails.

at1 moe verify house.at1w
#  { "experts": 8, "tensors_checked": 96,
#    "container_hash": "OK", "signature": "OK",
#    "VERDICT": "PASS" }

Why a verified vault

ApproachOne expert residentVerified weightsSigned routing
Load every expert into RAM
simple, but you pay for all N experts resident even though one answers
Serve each model separately
less RAM per call, but N deployments, no shared vault, no provable routing
partial
mmap a merged checkpoint
small working set, but no integrity, no signed router, no per-expert addressing
AT-1 Mixture-of-Specialists
one verified vault, one expert resident, signed + provable routing

The specialists are public open-source models; what AT-1 adds is the verified, addressable vault they live in, the signed router, and the streaming runtime — the byte-identical guarantee is the point, because the weights served are literally the same weights, just defer-loaded and hash-checked.

Built for

Multi-domain assistants (a coder, a math expert, a chat model in one vault) · on-device or cost-bounded serving where RAM is the constraint · regulated deployments that must prove which model answered a given request.

Verifying a container needs no account; packing and serving experts are metered against the connected account.