Mai-iaM/BeyondSight-aamt

BeyondSight models in .aamt

Five models packed in the .aamt file format, run on your own GPU by the aamt runtime:

npx aamt                  # InnerSight: chatting in about a minute
npx aamt run gaian        # the general 8B model
npx aamt run flow3        # Flow 3, 27B: the deep model
npx aamt bench gaian      # measure it on your machine and get a shareable card

model	file	size	parameters	weights
InnerSight	InnerSight.aamt	0.72 GB	1.7B	ternary
BeyondSight NightOwl	BeyondSight-NightOwl.aamt	2.28 GB	8B	1-bit
BeyondSight GaIaN	BeyondSight-GaIaN.aamt	3.23 GB	8B	ternary
InnerSight 2 4B (text)	InnerSight-2-4B.aamt	1.75 GB	4B	ternary
Flow 3	Flow-3.aamt	10.50 GB	27B	ternary

InnerSight 2 4B and Flow 3 are Prism ML's Ternary Bonsai 2 (4B and 27B), repacked losslessly from their own GGUF releases: the same ternary values, at a finer scale group. Both are hybrid-attention models (gated DeltaNet plus full attention). On the aamt runtime they agree with Prism ML's own MLX runtime on 99.3% and 100% of next-token predictions over the same text (the 4B's one disagreement is a near-tie). The 4B release is a vision model; this file is its text half only, and nothing here takes an image. They read the bonsai-tokenizer/ folder.

Weights are bit-packed so a GPU shader reads one 32-bit word and unpacks 16 ternary or 32 one-bit weights in registers. The format is specified at github.com/wwhitehead/aamt.

Provenance

• catalog.json lists every file with its size and SHA-384; catalog.sig is its Ed25519 signature by AsAManThinks Research. The aamt runtime verifies both before it loads a file, and npx aamt verify <file> checks any copy.
• The models carry a keyed watermark in their weight scales. It does not change their behaviour: on each model, the watermarked and original files agree on 100% of next-token predictions and produce identical greedy output on the runtime. Removing it breaches the licence.

Licence

BeyondSight Model License 1.0 (LICENSE.md): free for personal use, non-commercial research, a 60-day commercial evaluation, and small organisations (under US $1M revenue, US $5M raised and 25 people) after free registration. Everything else — production use, paid products, hosting, OEM, redistribution, training other models — needs a commercial licence: licensing@asamanthinks.com.

Derived from Qwen3 (Alibaba Cloud, Apache-2.0); see NOTICE and LICENSE-QWEN3-APACHE-2.0.txt. InnerSight 2 4B and Flow 3 are derived from Ternary Bonsai 2 by Prism ML (Apache-2.0), built on Qwen3.5-4B and Qwen3.8-27B; created using Bonsai by Prism ML. See NOTICE and LICENSE-BONSAI-2-APACHE-2.0.txt.