clemsail/micro-kiki-v35b

micro-kiki

35-domain expert model built on Qwen3.5-35B-A3B (MoE, 256 experts, 3B active/token) with LoRA adapters and a cognitive layer (memory palace + negotiator + anti-bias).

Model Description

micro-kiki is a multi-domain language model designed for technical applications spanning electronics, firmware, CAD, manufacturing, and general-purpose conversation. It uses a router-based architecture that selects up to 4 domain-specific LoRA stacks per request.

Property	Value
Base model	Qwen3.5-35B-A3B
Architecture	MoE (256 experts, 3B active/token)
Adapter	LoRA rank 16 (q/k/v/o projections)
Domains	35
Max active stacks	4
Context length	262,144 tokens
Quantization	Q4_K_M (inference), BF16 (training)
License	Apache 2.0

Architecture

                         +-------------------+
                         |   Domain Router   |
                         | (classifier, top4)|
                         +--------+----------+
                                  |
              +----------+--------+--------+----------+
              |          |                 |          |
         +----v----+ +---v---+       +----v----+ +---v---+
         | Stack 1 | |Stack 2|  ...  |Stack 34 | |Stack35|
         | chat-fr | |python |       |ml-train | |securi.|
         +---------+ +-------+       +---------+ +-------+
              |          |                 |          |
              +----------+--------+--------+----------+
                                  |
                         +--------v----------+
                         |    Negotiator     |
                         | CAMP + Catfish    |
                         +--------+----------+
                                  |
                         +--------v----------+
                         |    Anti-Bias      |
                         | KnowBias + RBD   |
                         +--------+----------+
                                  |
                         +--------v----------+
                         |   Aeon Memory     |
                         | Atlas + Trace     |
                         +-------------------+

Intended Use

• French/English conversational AI with domain expertise
• Code generation (Python, C/C++, Rust, TypeScript, embedded firmware)
• Electronics design (KiCad DSL, schematic review, component selection, SPICE)
• Manufacturing (process optimization, quality control)
• Multi-domain routing with cognitive arbitration

Limitations

• Not designed for medical, legal, or financial advice
• Optimized for technical domains; general knowledge may be weaker than base model
• Requires Q4_K_M or higher quantization; quality degrades below Q4
• Maximum 4 concurrent LoRA stacks; performance varies with stack combinations
• Memory (Aeon) requires external backends (Qdrant/Neo4j) for production use

Training Data — V3 (489K examples, 35 domains)

Sources

Source	Examples	Description
Claude CLI sessions	50,116	Real user-tool interactions extracted from 5 machines (GrosMac, kxkm-ai, Studio, Tower, CILS)
Codex/Copilot sessions	2,529	OpenAI Codex + GitHub Copilot sessions extracted from 4 machines
HuggingFace datasets	364,045	19 open datasets (see below)
Opus teacher distillation	—	chat-fr, reasoning domains
Original curated	—	32 domain seed datasets

HuggingFace Datasets

Dataset	Examples	License
CodeFeedback-Filtered-Instruction	157,000	Apache 2.0
French-Alpaca-Instruct-110K	110,000	Apache 2.0
Electronics StackExchange	95,000	CC-BY-SA-3.0
CJJones/LLM_EE_Educational_Synthetic_Dialog	50,000	CC-BY-NC-SA-4.0
MuratKomurcu/stm32-hal-dataset	29,700	MIT
redcathode/thingiverse-openscad	7,400	—
ThomasTheMaker/OpenSCAD	4,900	—
STEM-AI-mtl/Electrical-engineering	1,100	—
JITX open-components-database	151	—
Vrindarani/netlistgen	106	—

35 Domains

Group	Domains
Conversation	chat-fr, reasoning
Code	python, typescript, cpp, rust, html-css, shell, sql, yaml-json, lua-upy
Infrastructure	docker, devops, llm-orch, llm-ops (NEW), ml-training (NEW)
Electronics	kicad-dsl, kicad-pcb, spice, electronics, components (NEW), power, emc, dsp
Hardware	embedded, stm32, iot, platformio
CAD	freecad
Web	web-frontend, web-backend
Other	music-audio, math, security

Changes from V2: 3 new domains (components, llm-ops, ml-training). spice-sim merged into spice. stm32 is a sub-category of embedded.

New Domain: components

57K Q&A about electronic component specs, datasheets, sourcing, BOM, and cross-reference. Sources: Electronics StackExchange (filtered by component tags) + JITX open-components-database.

Training — V3

Property	Value
Base model	Qwen3.5-4B
Adapter	MoE-LoRA: 4 experts/projection, rank 16, top-2 routing
Null-space projection	ENABLED (prevents catastrophic forgetting between stacks)
Curriculum	Sequential, 35 stacks trained in order
Platform (MLX)	Mac Studio M3 Ultra 512 GB
Platform (CUDA)	kxkm-ai RTX 4090 24 GB

Evaluation

Metric	Value
Router accuracy (35-class)	[PENDING]
Forgetting check (angle)	[PENDING]
Perplexity (base)	[PENDING]
Perplexity (debiased)	[PENDING]
Aeon recall@1	[PENDING]
Aeon recall@5	[PENDING]
Aeon recall@10	[PENDING]
Anti-bias flag rate	[PENDING]
Average inference latency	[PENDING]

Hardware Requirements

Setup	RAM/VRAM	Use
Mac Studio M3 Ultra	512 GB unified	Training (BF16 LoRA) + serving (MLX)
RTX 4090	24 GB VRAM	Q4 inference (vLLM)
Apple Silicon 32 GB+	32 GB unified	Q4_K_M inference (MLX/llama.cpp)

Citation

@misc{micro-kiki-2026,
  title={micro-kiki: Multi-Domain Expert Model with Cognitive Layer},
  author={L'Electron Rare},
  year={2026},
  url={https://huggingface.co/electron-rare/micro-kiki}
}