kinetic-eqlm-46m-compute-matched

The flagship artifact of the Kinetic AI programme: a weight-tied transformer block applied twelve times, at the same width as a conventional 12-layer baseline, so that one iteration costs exactly one layer and compute is equal by construction.

The measured claim

	this model	explicit 12-layer baseline
parameters	45.8M	123.8M
resident weights (bf16)	183 MB	496 MB
compute per token	84.9M units	84.9M units (equal)
BLiMP ratio, 3 seeds	0.958 ± 0.017	1.000
BLiMP ratio, 31 phenomena	0.954	1.000

At identical arithmetic it delivers roughly 96% of the baseline's quality with 2.70 times fewer parameters (12 times fewer in the blocks themselves). One checkpoint serves every budget: at solver depths 4, 8 and 12 quality degrades gracefully (0.93 of baseline at half depth), which no fixed stack offers.

What is honestly NOT claimed

It is not better than the baseline on quality, and the gap (≈4%) did not close under depth modulation (that made it worse — the tying works because repetition of ONE map contracts to a fixed point). The memory saving is in weights, not activations: at batch one the Anderson solver's history makes activation peak 2.3× worse than the baseline. Scaling this recipe to open-weight-class models requires a pretraining budget: converting pretrained models fails at any gentleness (initial damage 64–270× base perplexity), and cheap logit distillation into the from-scratch student failed its pre-registered gate (−2.2% against a required +15%).

Files and formats

model.safetensors — exact, weight tying preserved, zero overhead. model_depth12.onnx — fixed 12-iteration graph with the block's tying preserved as shared initializers; the embedding/head tie is folded into two copies by the exporter, so the file is 337 MB rather than 183 MB. GGUF is deliberately not provided: llama.cpp cannot express weight sharing, and the required unrolling produces a file 4.9× larger than the baseline this model saves against.

Provenance

Findings F44–F53 in https://github.com/SharathSPhD/game-llm — every number traces to a results file with config hash and commit; the record includes the refuted attempts at the same evidentiary standard as the successes.