looped transformer · 16 × 3 = depth 48
Byrne‑100M UltraX‑MC
113.9M SpikeWhaleLM — MLA attention, fractal RoPE, depth attention, Hyper‑Connections, Engram n‑gram memory, MemoryCache branch, HRM refinement, MTP speculative decoding. Served by the package’s own engine, verified cache‑exact against a full recompute.
a small in-progress checkpoint: it holds a conversation and
repeats facts placed in its context, but its world knowledge is weak
16 512
0 1.5
0 100
1 2
temp 0 is greedy — deterministic, and it lands on generic attractors. Use it to compare runs, not to judge quality.
engine: spike_infer · 48 KV slots · cache-exact vs full recompute (max |Δlogit| 2.8e-05)