---
license: apache-2.0
base_model: Qwen/Qwen3.5-35B-A3B
base_model_relation: adapter
library_name: mlx
pipeline_tag: text-generation
language:
  - en
  - fr
tags:
  - lora
  - mlx
  - mlx-lm
  - qwen3.5
  - moe
  - spice
  - ngspice
  - electronics
  - micro-kiki
  - hypneum-lab
---

# micro-kiki stack-spice — SPICE-simulation LoRA for Qwen3.5-35B-A3B

One standard LoRA adapter (`adapters.safetensors`, 2.05 GB) trained with `mlx_lm` on
`Qwen/Qwen3.5-35B-A3B` for SPICE / ngspice / LTspice netlist and simulation assistance.
(The previous card called it a "merged adapter"; the shipped `adapter_config.json` shows
a plain LoRA run.)

## Recipe (from `adapter_config.json`)

| Parameter | Value |
|---|---|
| Base model | `Qwen/Qwen3.5-35B-A3B` (local `models/qwen3.5-35b-a3b`, BF16) |
| Framework | `mlx_lm.lora`, `fine_tune_type: lora` |
| LoRA rank / alpha / scale | 16 / 32 / **2.0** (= alpha/rank; unlike V4 which used scale 20) |
| Dropout | 0.0 |
| Target layers | 16 |
| Iterations | 11 542 |
| Batch size | 1 (`grad_accumulation_steps` 1) |
| Learning rate | 5e-5, no schedule, optimizer `adam` |
| Max sequence length | 2048 |
| Gradient checkpointing | true |
| Seed | 0 |
| Data | `outputs/stacks/stack-spice` (micro-kiki `spice` domain; example counts not recorded) |
| Adapter size | 2 051 644 214 bytes (16 layers × ~128 MB) |

## Training host & cost

Mac Studio M3 Ultra 512 GB, MLX BF16. Peak memory, wall-clock, energy: not recorded
(no log shipped). Host powered off since 2026-10.

## Evaluation

Not recorded for this adapter. Related evidence: the project README reports that the base
model is already strong on SPICE/electronics and that only 3 of 10 probed domains improved
over base; in the June 2026 iact-bench run on the sibling Qwen3.6 ailiance adapters, no
adapter beat the base except `html-css`, and a SPICE adapter was removed from production
routing for degrading numeric (RC) calculations. Validate on your own netlists before use.

## Limitations

Research artefact; may produce syntactically plausible but electrically wrong netlists;
not for safety-critical design sign-off.

## Citation

micro-kiki concept DOI **10.5281/zenodo.19666759**.

## Provenance

<https://github.com/hypneum-lab/micro-kiki> · <https://github.com/ailiance/ailiance-mac-tuner>
· base <https://huggingface.co/Qwen/Qwen3.5-35B-A3B>

## EU AI Act transparency

LoRA adapter weights only, base unchanged; provider Clément Saillant / L'Electron Rare for
Hypneum Lab; limited risk, Article 50; contact clement@saillant.cc. Outputs may be
inaccurate, biased or fabricated.