Instructions to use clemsail/micro-kiki-v3 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use clemsail/micro-kiki-v3 with MLX:

# Make sure mlx-lm is installed
# pip install --upgrade mlx-lm

# Generate text with mlx-lm
from mlx_lm import load, generate

model, tokenizer = load("clemsail/micro-kiki-v3")

prompt = "Write a story about Einstein"
messages = [{"role": "user", "content": prompt}]
prompt = tokenizer.apply_chat_template(
    messages, add_generation_prompt=True
)

text = generate(model, tokenizer, prompt=prompt, verbose=True)

Notebooks
Google Colab
Kaggle
Local Apps Settings

llama.cpp

How to use clemsail/micro-kiki-v3 with llama.cpp:

Install (macOS, Linux)

curl -LsSf https://llama.app/install.sh | sh
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf clemsail/micro-kiki-v3:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf clemsail/micro-kiki-v3:Q4_K_M

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf clemsail/micro-kiki-v3:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf clemsail/micro-kiki-v3:Q4_K_M

Use pre-built binary

# Download pre-built binary from:
# https://github.com/ggerganov/llama.cpp/releases
# Start a local OpenAI-compatible server with a web UI:
./llama-server -hf clemsail/micro-kiki-v3:Q4_K_M
# Run inference directly in the terminal:
./llama-cli -hf clemsail/micro-kiki-v3:Q4_K_M

Build from source code

git clone https://github.com/ggerganov/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-server llama-cli
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama-server -hf clemsail/micro-kiki-v3:Q4_K_M
# Run inference directly in the terminal:
./build/bin/llama-cli -hf clemsail/micro-kiki-v3:Q4_K_M

Use Docker

docker model run hf.co/clemsail/micro-kiki-v3:Q4_K_M

LM Studio
Jan

vLLM

How to use clemsail/micro-kiki-v3 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "clemsail/micro-kiki-v3"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "clemsail/micro-kiki-v3",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/clemsail/micro-kiki-v3:Q4_K_M

Ollama
How to use clemsail/micro-kiki-v3 with Ollama:
```
ollama run hf.co/clemsail/micro-kiki-v3:Q4_K_M
```
Unsloth Desktop

How to use clemsail/micro-kiki-v3 with Pi:

Start the MLX server

# Install MLX LM:
uv tool install mlx-lm
# Start a local OpenAI-compatible server:
mlx_lm.server --model "clemsail/micro-kiki-v3"

Configure the model in Pi

# Install Pi:
npm install -g @earendil-works/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "mlx-lm": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "clemsail/micro-kiki-v3"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

MLX LM

How to use clemsail/micro-kiki-v3 with MLX LM:

Generate or start a chat session

# Install MLX LM
uv tool install mlx-lm
# Interactive chat REPL
mlx_lm.chat --model "clemsail/micro-kiki-v3"

Run an OpenAI-compatible server

# Install MLX LM
uv tool install mlx-lm
# Start the server
mlx_lm.server --model "clemsail/micro-kiki-v3"
# Calling the OpenAI-compatible server with curl
curl -X POST "http://localhost:8000/v1/chat/completions" \
   -H "Content-Type: application/json" \
   --data '{
     "model": "clemsail/micro-kiki-v3",
     "messages": [
       {"role": "user", "content": "Hello"}
     ]
   }'

Docker Model Runner
How to use clemsail/micro-kiki-v3 with Docker Model Runner:
```
docker model run hf.co/clemsail/micro-kiki-v3:Q4_K_M
```

Lemonade

How to use clemsail/micro-kiki-v3 with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull clemsail/micro-kiki-v3:Q4_K_M

Run and chat with the model

lemonade run user.micro-kiki-v3-Q4_K_M

List all available models

lemonade list

Hermes Agent

How to use clemsail/micro-kiki-v3 with Hermes Agent:

Start the MLX server

# Install MLX LM:
uv tool install mlx-lm
# Start a local OpenAI-compatible server:
mlx_lm.server --model "clemsail/micro-kiki-v3"

Configure Hermes

# Install Hermes:
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
hermes setup
# Point Hermes at the local server:
hermes config set model.provider custom
hermes config set model.base_url http://127.0.0.1:8080/v1
hermes config set model.default clemsail/micro-kiki-v3

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use clemsail/micro-kiki-v3 with OpenClaw:

Start the MLX server

# Install MLX LM:
uv tool install mlx-lm
# Start a local OpenAI-compatible server:
mlx_lm.server --model "clemsail/micro-kiki-v3"

Configure OpenClaw

# Install OpenClaw:
npm install -g openclaw@latest
# Register the local server and set it as the default model:
openclaw onboard --non-interactive --mode local \
  --auth-choice custom-api-key \
  --custom-base-url http://127.0.0.1:8080/v1 \
  --custom-model-id "clemsail/micro-kiki-v3" \
  --custom-provider-id mlx-lm \
  --custom-compatibility openai \
  --custom-text-input \
  --accept-risk \
  --skip-health

Run OpenClaw

openclaw agent --local --agent main --message "Hello from Hugging Face"

micro-kiki v3 — pre-pivot MoE-LoRA stacks on Qwen3.5-4B (DEPRECATED: untrained weights)

Do not use. The 35 adapters in this repository are the pre-pivot stacks-v3-r16 lot. An audit on 2026-04-19 (docs/research/2026-04-19-prepivot-moe-lora-audit.md in hypneum-lab/micro-kiki) found that every lora_B tensor is exactly zero on all adapters tested from this lot: the LoRA delta A·B is 0, the adapters change nothing. The custom MoE-LoRA training path never back-propagated into B. The training status table of ailiance/ailiance-mac-tuner lists this lot as DEAD. The pivot to standard LoRA (micro-kiki-v35b, then micro-kiki-v4-sota) replaced it.

Corrections vs. the previous card. The earlier card described a 35-domain expert model on Qwen3.5-35B-A3B with a cognitive layer. The files say otherwise:

base model is Qwen3.5-4B (dense), not 35B-A3B: the shipped micro-kiki-v3-Q4_K_M.gguf has 4 205 751 296 parameters (architecture qwen35);
the adapters are custom MoE-LoRA (4 experts per projection, rank 16, router heads), 35 × 481 680 416 bytes, with stack_meta.json recording train_loss: 0.0 for every domain (500 steps, 2026-04-17);
because the LoRA B matrices are zero, the fused GGUF is equivalent to the quantized base Qwen3.5-4B;
all evaluation fields of the earlier card were [PENDING]; none was ever filled.

What is kept and why

The repository is kept (deprecated) as the reproducible artefact of the training-path bug documented in the audit; results/forgetting-matrix-prepivot.json in the GitHub repo holds the 0° pairwise matrix computed from these files. Weights may be removed in a later cleanup; this card and adapters/*/stack_meta.json will remain.

Recipe (as configured, not effective)

Parameter	Value
Base model	`Qwen/Qwen3.5-4B`
Adapter type	custom MoE-LoRA, 4 experts/projection, rank 16, top-2 routing, null-space projection enabled
Steps	500 per stack, sequential curriculum over 35 domains
Framework / host	MLX BF16 on Mac Studio M3 Ultra 512 GB (also CUDA experiments on an RTX 4090)
Data	micro-kiki v3 corpus, 489 K examples / 35 domains (see GitHub)
Effective result	`lora_B = 0` everywhere → no learned delta

Evaluation

None valid: any evaluation of these adapters measures the base model.

Citation

micro-kiki concept DOI 10.5281/zenodo.19666759. Successor: hypneum-lab/micro-kiki-v4-sota.

Provenance

https://github.com/hypneum-lab/micro-kiki (scripts/legacy/, audit) · https://github.com/ailiance/ailiance-mac-tuner (status table).

EU AI Act transparency

Adapter files with zero effective delta; base Qwen/Qwen3.5-4B unchanged; provider Clément Saillant / L'Electron Rare; contact clement@saillant.cc.

Downloads last month: 65

MLX

Hardware compatibility

Quantized

GGUF

Model size

4B params

Architecture

qwen35

Hardware compatibility

4-bit

Model tree for clemsail/micro-kiki-v3

Base model

Qwen/Qwen3.5-4B-Base

Finetuned

Qwen/Qwen3.5-4B

Adapter

(834)

this model