spikingkiki-35b-a3b-v4

clemsail/spikingkiki-35b-a3b-v4

spikingkiki-35b-a3b-v4 is a machine learning model from clemsail. Use it for the machine learning task on the model card, and read the license before you ship it in a product. The card lists the license as apache-2.0.

Deprecated / incomplete. This repository contains 9 988 per-module .npz files (19.9 GB) out of the 31 070 modules (≈58-70 GB) produced by the conversion. Entire expert and attention modules are missing (e.g. layer 0 h…

Downloads · 30 days

Access

Public

Updated Oct 25, 2026

Repo size

19.9 GB

Likes

Public

Hugging Face

Repo makeup

Click a slice to open those files.

.npz19.9 GB · 100%

At a glance

License: apache-2.0
Access: Public
Created: Apr 22, 2026
Updated: Oct 25, 2026
SHA: ec9c52de

Base models

Qwen/Qwen3.6-35B-A3B

Model card

From the Hugging Face model README

SpikingKiki-35B-A3B-V4 — partial upload of a rate-coded SNN conversion (deprecated)

Deprecated / incomplete. This repository contains 9 988 per-module .npz files (19.9 GB) out of the 31 070 modules (≈58-70 GB) produced by the conversion. Entire expert and attention modules are missing (e.g. layer 0 has 168 of 256 experts; self_attn.o_proj is present for 1 of 40 layers). It cannot be loaded as a model. The complete artefact exists only on the (now powered-off) build host.

What the full artefact was: a lossless ANN-to-SNN (LAS) rate-coded conversion of Qwen/Qwen3.6-35B-A3B, each linear module replaced by an integrate-and-fire surrogate (T = 128 timesteps, nominal threshold 0.0625, tau 1.0; per-module thresholds recorded in lif_metadata.json). Not a fused model and not sharded weights: one .npz per module.

Recipe

Item	Value
Source	`Qwen/Qwen3.6-35B-A3B` BF16 weights
Script	`scripts/convert_spikingkiki_35b.py` in `hypneum-lab/micro-kiki`
Parameters	T = 128, threshold 0.0625 (nominal), tau 1.0
Output	31 070 module files; 40/40 transformer blocks (Phase C)
Validation	99.5 % sane modules on N = 200 (Phase E)
Uploaded here	9 988 modules, `lif_metadata.json`, `.convert_state.json`

Build host & cost

Mac Studio M3 Ultra 512 GB; 1 h 08 wall-clock with 4-worker block-range parallelism (11 086 s reported in the tuner repo). Energy: not recorded. Host off since 2026-10.

Evaluation

No functional evaluation (no quantisation, smoke test or energy benchmark was run after the conversion). The 0.032 mJ/token figure in the project README is a theoretical estimate, not a measurement.

Status / what to do

Either complete the upload from the original host, or treat this repository as an archive of metadata only. Downstream consumer hypneum-lab/dream-of-kiki (DREAM_MICRO_KIKI_REAL=1) cannot use the partial set.

Citation

micro-kiki concept DOI 10.5281/zenodo.19666759.

Provenance

https://github.com/hypneum-lab/micro-kiki (docs/training/spikingkiki-conversion-runbook.md, results/spikingkiki-35b-analysis.json) · https://github.com/ailiance/ailiance-mac-tuner

EU AI Act transparency

Derived (converted) weights of Qwen/Qwen3.6-35B-A3B, no additional training data; provider Clément Saillant / L'Electron Rare for Hypneum Lab; research-only; contact [email protected].

License: apache-2.0
Created: Apr 22, 2026
Updated: Oct 25, 2026