Configurations
Each method in the table has an admitted configuration, configs/<family>/<method>.json, and a cost proof in the current
proof set. The table is generated from those files: physical blocks, FFN width, independent parameters, audited training
GFLOPs per input token and the deviation from the SSSL anchor (4.67239832544785). Parameters are outcomes, not controls.
| Family | Method | Blocks | FFN | Parameters | GFLOPs / token | vs SSSL | Stages |
|---|---|---|---|---|---|---|---|
| baseline | dense_full |
23 | 6,144 | 1,355,810,568 | 4.674662 | +0.048% | joint |
| looped | deeploop |
12 | 6,144 | 742,392,722 | 4.660529 | -0.254% | joint |
| looped | fixed_recursive |
15 | 6,144 | 927,990,968 | 4.663054 | -0.200% | joint |
| looped | huginn |
14 | 6,528 | 870,581,286 | 4.649327 | -0.494% | joint |
| looped | ouro |
9 | 6,144 | 607,127,805 | 4.658210 | -0.304% | joint, gate_only |
| looped | ut_act |
12 | 6,144 | 742,394,259 | 4.660085 | -0.264% | joint |
| sparse | dsa |
22 | 6,400 | 1,316,236,406 | 4.650934 | -0.459% | burnin, joint |
| sparse | moba |
25 | 6,144 | 1,462,765,468 | 4.673521 | +0.024% | joint |
| sparse | nsa |
23 | 6,144 | 1,381,293,832 | 4.638686 | -0.722% | joint |
| sparse | qsa |
23 | 6,144 | 1,378,420,488 | 4.628037 | -0.949% | burnin, joint |
| sparse | seer |
23 | 6,272 | 1,387,464,456 | 4.659564 | -0.275% | burnin, joint |
| linear | deltanet |
25 | 6,144 | 1,463,690,268 | 4.674979 | +0.055% | joint |
| linear | elu_norm |
25 | 6,144 | 1,462,768,668 | 4.651242 | -0.453% | joint |
| linear | elu_qknorm |
25 | 6,144 | 1,462,768,668 | 4.652243 | -0.431% | joint |
| linear | gdn |
23 | 6,144 | 1,411,349,680 | 4.687197 | +0.317% | joint |
| linear | gla |
23 | 6,144 | 1,411,207,816 | 4.642048 | -0.650% | joint |
| linear | kda |
23 | 6,144 | 1,412,126,344 | 4.692443 | +0.429% | joint |
| linear | retnet |
23 | 6,144 | 1,410,077,320 | 4.630060 | -0.906% | joint |
| residual | attnres |
23 | 6,144 | 1,355,882,760 | 4.679857 | +0.160% | joint |
| residual | hc |
23 | 6,144 | 1,356,235,700 | 4.698483 | +0.558% | joint |
| residual | mhc |
23 | 6,016 | 1,353,550,818 | 4.674363 | +0.042% | joint |
| residual | residual_standard |
23 | 6,144 | 1,355,810,568 | 4.674298 | +0.041% | joint |
| hybrid | deltanet_hybrid |
24 | 6,400 | 1,403,810,314 | 4.675326 | +0.063% | joint |
| hybrid | gdn_hybrid |
24 | 5,760 | 1,399,275,962 | 4.703800 | +0.672% | joint |
| hybrid | gdn_hybrid_swa |
22 | 6,528 | 1,329,681,662 | 4.647087 | -0.542% | joint |
| hybrid | kda_hybrid |
24 | 6,016 | 793,542,986 | 4.692309 | +0.426% | joint |
| hybrid | moba_hybrid |
25 | 6,016 | 1,452,935,068 | 4.659989 | -0.266% | joint |
| hybrid | qsa_hybrid |
24 | 5,760 | 1,405,174,202 | 4.691637 | +0.412% | burnin, joint |
| others | dense (SSSL) |
24 | 6,144 | 1,384,122,122 | 4.672398 | +0.000% | joint |
Blocks count physical layers; looped methods reuse them across loops, so their unfolded depth is larger (see the source notes). Multi-stage methods weight each stage by its updates: DSA, QSA, Seer and the QSA hybrid run 512 dense burn-in updates before sparse training; Ouro trains its exit gate alone for the final 1,024 updates; Huginn samples a depth per update from a fixed seed, and its proof enumerates all 46,875 samples.
Reading a configuration
Section titled “Reading a configuration”A configuration holds the model fields (model), the protocol it is admitted under (protocol, with its fingerprint)
and the calibration record (calibration: the audited FLOPs per token and their deviation, the method’s source
fingerprint, the PyTorch build of the audit and the path of the cost proof). Configurations are written by
python -m nexttransformer.protocol.calibrate and checked by admit_config; never edit them by hand.
The historical native track (SSSL, MoR, LoopFormer, Circle Loop) keeps its own configurations under configs/native/,
with the historical matrix-FLOPs accounting.