Instructions to use miguelamendez/mica-spark-x25-4b with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use miguelamendez/mica-spark-x25-4b with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="miguelamendez/mica-spark-x25-4b")

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoModel
model = AutoModel.from_pretrained("miguelamendez/mica-spark-x25-4b", device_map="auto")

MLX

How to use miguelamendez/mica-spark-x25-4b 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("miguelamendez/mica-spark-x25-4b")

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 miguelamendez/mica-spark-x25-4b 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 miguelamendez/mica-spark-x25-4b:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf miguelamendez/mica-spark-x25-4b:Q4_K_M

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf miguelamendez/mica-spark-x25-4b:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf miguelamendez/mica-spark-x25-4b: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 miguelamendez/mica-spark-x25-4b:Q4_K_M
# Run inference directly in the terminal:
./llama-cli -hf miguelamendez/mica-spark-x25-4b: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 miguelamendez/mica-spark-x25-4b:Q4_K_M
# Run inference directly in the terminal:
./build/bin/llama-cli -hf miguelamendez/mica-spark-x25-4b:Q4_K_M

Use Docker

docker model run hf.co/miguelamendez/mica-spark-x25-4b:Q4_K_M

LM Studio
Jan

vLLM

How to use miguelamendez/mica-spark-x25-4b with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "miguelamendez/mica-spark-x25-4b"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "miguelamendez/mica-spark-x25-4b",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/miguelamendez/mica-spark-x25-4b:Q4_K_M

SGLang

How to use miguelamendez/mica-spark-x25-4b with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "miguelamendez/mica-spark-x25-4b" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "miguelamendez/mica-spark-x25-4b",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "miguelamendez/mica-spark-x25-4b" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "miguelamendez/mica-spark-x25-4b",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Ollama
How to use miguelamendez/mica-spark-x25-4b with Ollama:
```
ollama run hf.co/miguelamendez/mica-spark-x25-4b:Q4_K_M
```
Unsloth Desktop

How to use miguelamendez/mica-spark-x25-4b with Pi:

Start the MLX server

# Install MLX LM:
uv tool install mlx-lm
# Start a local OpenAI-compatible server:
mlx_lm.server --model "miguelamendez/mica-spark-x25-4b"

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": "miguelamendez/mica-spark-x25-4b"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

MLX LM

How to use miguelamendez/mica-spark-x25-4b with MLX LM:

Generate or start a chat session

# Install MLX LM
uv tool install mlx-lm
# Interactive chat REPL
mlx_lm.chat --model "miguelamendez/mica-spark-x25-4b"

Run an OpenAI-compatible server

# Install MLX LM
uv tool install mlx-lm
# Start the server
mlx_lm.server --model "miguelamendez/mica-spark-x25-4b"
# Calling the OpenAI-compatible server with curl
curl -X POST "http://localhost:8000/v1/chat/completions" \
   -H "Content-Type: application/json" \
   --data '{
     "model": "miguelamendez/mica-spark-x25-4b",
     "messages": [
       {"role": "user", "content": "Hello"}
     ]
   }'

Docker Model Runner
How to use miguelamendez/mica-spark-x25-4b with Docker Model Runner:
```
docker model run hf.co/miguelamendez/mica-spark-x25-4b:Q4_K_M
```

Lemonade

How to use miguelamendez/mica-spark-x25-4b with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull miguelamendez/mica-spark-x25-4b:Q4_K_M

Run and chat with the model

lemonade run user.mica-spark-x25-4b-Q4_K_M

List all available models

lemonade list

Hermes Agent

How to use miguelamendez/mica-spark-x25-4b 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 "miguelamendez/mica-spark-x25-4b"

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 miguelamendez/mica-spark-x25-4b

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use miguelamendez/mica-spark-x25-4b with OpenClaw:

Start the MLX server

# Install MLX LM:
uv tool install mlx-lm
# Start a local OpenAI-compatible server:
mlx_lm.server --model "miguelamendez/mica-spark-x25-4b"

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 "miguelamendez/mica-spark-x25-4b" \
  --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"

Spark-X2.5-4B — Mica Q4/Q8

Mica runtime artifacts for XHToken/Spark-X2.5-4B, pinned to revision 0bcb35678590218655dff3765b9e61c83b35e9c4. The source and derivatives are Apache-2.0 licensed.

The runtime configuration, conversion policy, and validation evidence are maintained in the Mica Server repository. Validated artifacts are published together in the miguelamendez/mica-spark-x25-4b repository, with manifests recording their source, size, and digest.

Model facts and provenance

Field	Value	Meaning
Parameter class	4B	Upstream model name; an exact parameter total is not published in the card
Architectural context	1,048,576 tokens	Declared by the pinned configuration; input plus output
Training context	sequences extending to 1M tokens	Upstream reports a dedicated long-context stage
Pretraining	approximately 20T tokens	Upstream disclosure
Maximum input alone	not disclosed	Do not infer it from the total context
Maximum generated output alone	not disclosed	The upstream 131,072-token example is not treated as a hard model limit
Thinking	supported, enabled by default	The chat template also permits `enable_thinking=false`
Thinking effort levels	not disclosed	Mica uses token budgets, not invented low/medium/high semantics

The architecture combines one full-attention layer with three sliding-window layers. Upstream reports more than 200 languages and evaluates the model in thinking mode. Training data categories are described at a high level, but an itemized dataset inventory is not published.

Mica artifacts

The repository contains all validated formats:

mlx/q4/    MLX affine Q4, group size 64
mlx/q8/    MLX affine Q8, group size 64
gguf/      Q4_K_M and Q8_0
vllm/      not published; structural candidates failed the production gate

The intended portable Linux/NVIDIA comparison was AutoRound W4A16 group-128 versus GPTQ W4A16 group-128, followed by GPTQ W8A16 group-128. AutoRound 0.14.2 cannot calibrate this Spark revision's custom four-dimensional attention mask, so GPTQ is now the active Q4 path. The embedding and lm_head stay at source precision. Candidates use compressed-tensors and remain outside the serving catalog until the custom Spark architecture loads and passes real vLLM generation plus BF16 quality comparison. Blackwell-only formats are not part of this matrix.

Two-sample structural GPTQ candidates for both W4A16 and W8A16 now serialize, reload, and generate with batch size two under the 16-GiB guard. They are not production artifacts; Q4's much higher calibration error and the tiny local fixture make the configured 512-sample calibration and BF16 comparison mandatory.

Weight precision and KV-cache precision are independent. A runtime profile may, for example, use Q4 weights with a Q8 KV cache.

Mica validation boundary

On the Apple M4 host, MLX Q4 and Q8 passed real generation and server inference. Measured peak model memory was 2.403 GiB for Q4 and 4.453 GiB for Q8 in the recorded smoke cases. GGUF Q4_K_M and Q8_0 also passed real Metal inference.

These tests do not validate the upstream 1M-token claim locally. Until the long-context matrix passes, Mica profiles must use their own certified context limit and warn above the longest locally tested input. The smaller 1.7B sibling failed the lossless classic-speculation gate and is not an accepted drafter. No native MTP or DFlash checkpoint was found for this target. The Apple M4 vLLM-Metal probe failed during custom-config parsing before model load, and the architecture is not in the plugin's support table.

Validation records:

docs/validation/spark-x25-4b-mlx-q4.md
docs/validation/spark-x25-4b-mlx-q8.md
docs/validation/spark-x25-4b-mlx-classic-speculative.md
docs/validation/spark-vllm-quantization-macos.md
docs/validation/gguf-runtime-macos-metal.md

Profile validation requirements

A Spark profile must satisfy:

max_input_tokens + max_output_tokens <= 1,048,576
max_reasoning_tokens + min_visible_output_tokens <= max_output_tokens
active_batch_tokens <= configured KV/cache admission budget

Exceeding Mica's tested context is a warning in experimental mode and an error in certified mode. Exceeding the architectural total is always an error.

Downloads last month: 144

MLX

Hardware compatibility

Quantized

GGUF

Model size

4B params

Architecture

spark2_5

Hardware compatibility

4-bit

8-bit

Model tree for miguelamendez/mica-spark-x25-4b

Base model

XHToken/Spark-X2.5-4B-Base

Finetuned

XHToken/Spark-X2.5-4B

Quantized

(54)

this model

Collection including miguelamendez/mica-spark-x25-4b

Mica Local Text Models

Collection

Permissively licensed local text-generation artifacts validated by Mica. • 1 item • Updated 29 days ago