Model Registry

Models

Every entry was trained end-to-end on lab hardware. Registry cards state architecture, token budgets, and evaluation numbers exactly as measured — mid-training checkpoints included.

HELIX-200M

Training

Hybrid H-E-R stack · 28 layers · d_model 1024 · HarmonicStateLayer + ElasticAttentionGate + ResonantExpertRouting (MoE top-2 of 4)

Params: 209.05M
Tokens: 40B
Val PPL: 39.4

HELIX-120M

Planned

H-E-R stack · 15 layers · d_model 768 · 12 heads · MoE top-2 of 4

Params: 127M
Tokens: 2.5B
Val PPL: —

HELIX-56M

Released

H-E-R stack · 12 layers · d_model 512 · 8 heads · 4 experts

Params: 57M
Tokens: Chinchilla-matched
Val PPL: 48.6

SLM-10M

Released

Causal transformer · 12 layers · 256 dim · GQA 8Q/2KV · QK-norm · RoPE θ=100k · SwiGLU · weight tying

Params: 9.97M
Tokens: 25B
Val PPL: —

D-CSIL-1

Archived

GLA · 12 layers · 256 dim · state 64 · 4 heads · seq 256

Params: 55.98M total · 43.4M active
Tokens: Multi-epoch
Val PPL: 455.77

D-CSIL-2

Archived

SSM+GLA hybrid line · benchmark-harness generation

Params: ~100M class
Tokens: Trained
Val PPL: 58.04

D-CSIL-3

Training

Selective State Language Model (SSLM) · all-local 7-stage pipeline

Params: 235.68M
Tokens: 4.747B
Val PPL: —

SynapNet V3

Archived

O(L) SSM+GLA hybrid · 12 layers · 256 dim · state 64 · Hebbian/neuromodulation/sleep mechanisms

Params: 42.57M
Tokens: WikiText-103,
Val PPL: 66.03

HibbieFormer (recurrent core)

Archived

Recurrent bio-inspired core · gated state blocks (hidden 64) + proto-neurons with 4 dendritic branches · 8 layers · d128–d768 family

Params: 13.5M / 27M / 60M / 84.4M family · 81.09M cycle-200 checkpoint
Tokens: TinyStories
Val PPL: 10.2

FieldTransformer v1.4

Training

MultiScaleFieldLayer (k=3 + k=7 Laplacian) + causal attention · flux / Schwarzian / latent-diffusion / entropy regularizers

Params: 9.0M (v1.4) · 1.36M (v1.0–1.3 char-level)
Tokens: Tiny
Val PPL: —