Minds that run on your own GPU.
BeyondSight is a family of models packed in .aamt and run by the Vortex engine — from InnerSight, small enough for a browser tab, to deep thinking models. Below, two towns think with them, and their memories converge across tabs and devices as one file.
- records
- 0
- clock
- —
- .aamt
- —
- records
- 0
- clock
- —
- .aamt
- —
In the replay, only the citizens' words come from the recording; your tab recomputes every projection, recall, record, merge, pack and hash — and two tabs must end on the recording's final hash. Who walks where is scripted; what they think is InnerSight's.
Every decision has an address.
Most models are a folder of weights you take on trust. A BeyondSight model is one signed .aamt file, and the minds built on it think in coordinates. So you can check what you loaded, who made it, and why a mind remembered what it did.
One file, signed
A BeyondSight model is one .aamt file, the same format as the towns' memories above. The runtime checks an Ed25519-signed catalog with a pinned key, and each file's SHA-384, before it loads anything.
Weights that carry their maker
A keyed watermark lives in the weights' scales. On all three released models it is detected at z 733 to 1,600, survives re-baking and f16/bf16 rounding, and leaves greedy output identical to the original.
Every thought has an address
Each thought becomes a (T, E, R, A) coordinate: Thought, Emotion, Reason, Action. It routes to one cell of the .aamt tree, an Odu word like 5.6.13.1. Recall is a route you can read: which cell, which memories, and how far by Fisher–Rao distance.
Every route is journaled
libaamt's lineage journal writes 40 bytes per routing decision, bound to the file's content hash. Replay it later and every decision must land in the same cell, bit for bit. Change the file and the journal says which decisions it touched.
Audit this run
Start the towns above. Every recall a citizen makes is journaled here.
What you're watching
GaIaN designs two towns
BeyondSight GaIaN, 8B ternary, writes each town as validated JSON: the places, the people, their trades and homes. Rejected drafts are kept in the record.
InnerSight is every mind
Each citizen's thoughts come from InnerSight, 1.7B ternary in a 723 MB file — small enough to run in a browser tab on your own graphics card.
Each world is a replica
A thought becomes a record in its world's CRDT replica, stamped with a hybrid logical clock, and is packed into a real .aamt file. The file's hash is the replica's identity.
The worlds converge
Records cross between the worlds — inside this tab, between tabs, or between devices — and merge. Merging is commutative, associative and idempotent, so any order of arrival ends on the same file and the same hash.
What BeyondSight can do
Measured on an Apple M1 Max unless it says otherwise.
The minds on this page can think on your GPU through WebGPU. Nothing is sent to a server to think.
An 8B model in 2.28 GB, with a 2.5 s cold prefill — fast enough to voice a town's people in real time.
Both towns here were designed by GaIaN as structured JSON that a validator checks before the engine builds it.
The runtime keeps a model's reading: the follow-up answers in 307 ms instead of 30.4 s, and branches start 55× faster.
The runtime verifies the signed catalog before it loads a model. The watermark changes 0% of next-token predictions.
Base models, experts for code, mathematics and Unity, a speculative draft, a 27B low-memory backbone and deep thinking models.
The family
Twelve checkpoints, all in .aamt. Three are released on Hugging Face today, signed; the rest are part of the family and not yet released.
| Model | Size | Weights | Role | For |
|---|---|---|---|---|
InnerSight released · signed | 0.72 GB | 1.7B · ternary from Qwen3-1.7B | base | The small companion: runs in a browser tab on your own graphics card. It is the mind of every citizen on this page. 723 MB, cached after the first load |
NightOwl released · signed | 2.28 GB | 8B · 1-bit from Qwen3-8B | base | Real-time dialogue and fast characters: an 8B model in 2.28 GB. 36.2 tok/s decode, 2.5 s cold prefill (M1 Max) |
GaIaN released · signed | 3.23 GB | 8B · ternary from Qwen3-8B | base | The general model, and the best measured quality of the three released. It designed both towns on this page. 41.2 tok/s sustained decode (M1 Max) |
GaIaN Draft not yet released | 0.47 GB | — | draft | A speculative-decoding draft: it guesses tokens for a larger model to check in one pass. |
GaIaN Code Expert not yet released | 3.07 GB | — | expert | A GaIaN expert trained for code. |
GaIaN Math Expert not yet released | 3.07 GB | — | expert | A GaIaN expert trained for mathematics. |
GaIaN Unity & C# Expert not yet released | 3.07 GB | — | expert | A GaIaN expert trained for Unity and C#: game code, in the engine's own idiom. |
GaIaN Rebake Control not yet released | 3.07 GB | — | research control | A research control: the same bake without expert training, so each expert's gain can be measured against it. |
Nomad not yet released | 7.21 GB | 27B · 1-bit | base | A low-memory 27B agent backbone, and a draft model for the largest checkpoints. 10–12 tok/s (M1 Max) |
Ezmarelda not yet released | 7.21 GB | — | base | A base model, the twin of Nomad at the same size. |
Flow not yet released | 10.41 GB | — | base | A thinking model for deep work: 64 layers, 48 of them linear attention, so its memory per token stays small. 11.3 tok/s decode; KV cache 268 MB at 4,096 tokens (M1 Max) |
Flow 2 not yet released | 10.50 GB | — | base | For precise logic and coding agents, and the default deep slot in MaiiaM OS: the model commands, reflection and planning are routed to. |
Released models are derived from Qwen3 (Apache-2.0) and ship under the BeyondSight Model License 1.0: free for personal use, research and evaluation; commercial use needs a licence.
Run BeyondSight yourself
One command: InnerSight chatting in about a minute, then GaIaN, and a benchmark card for your own machine.
npx @asamanthinks/aamt