electron-rare/spikingkiki-122b-a10b

SpikingKiKi 122B (A10B)

Spiking Neural Network (SNN) variant of Qwen3.5 MoE 122B with ~10B active parameters per token, using Leaky Integrate-and-Fire (LIF) neurons in selected layers. Research artifact from Hypneum Lab.

• —Base architecture: Qwen3.5 MoE (122B total, ~10B active)
• —Modification: LIF spiking neurons in attention/FFN routing layers
• —Metadata: see lif_metadata.json for per-layer LIF parameters
• —Status: research-stage; not benchmarked end-to-end
• —Shards: 28 safetensors files (full weights, ~244 GB)

Files

• —model-*.safetensors (28 shards) — full weights
• —config.json — Qwen3.5 MoE architecture config
• —lif_metadata.json — LIF layer parameters
• —chat_template.jinja — standard Qwen3.5 chat template

Research context

Extends the SpikingKiki line (see electron-rare/spikingkiki-27b) to MoE architecture, exploring whether sparse activation routing composes synergistically with spiking neurons for further compute reduction. Related to the GENIAL framework on `hypneum-lab`.

Org context

This repo is part of electron-rare, the legacy research org. Production ailiance artifacts (gateway, mascarade LoRAs) live at `Ailiance-fr` (post-2026-05-11 carve-out).