Instructions to use XHToken/Spark-X2.6-4B-Base with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use XHToken/Spark-X2.6-4B-Base with Transformers:

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

pipe = pipeline("text-generation", model="XHToken/Spark-X2.6-4B-Base", trust_remote_code=True)
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained("XHToken/Spark-X2.6-4B-Base", trust_remote_code=True, device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use XHToken/Spark-X2.6-4B-Base with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "XHToken/Spark-X2.6-4B-Base"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "XHToken/Spark-X2.6-4B-Base",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/XHToken/Spark-X2.6-4B-Base

SGLang

How to use XHToken/Spark-X2.6-4B-Base 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 "XHToken/Spark-X2.6-4B-Base" \
    --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": "XHToken/Spark-X2.6-4B-Base",
		"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 "XHToken/Spark-X2.6-4B-Base" \
        --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": "XHToken/Spark-X2.6-4B-Base",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use XHToken/Spark-X2.6-4B-Base with Docker Model Runner:
```
docker model run hf.co/XHToken/Spark-X2.6-4B-Base
```

CMF build of Spark-X2.6-4B for cortiq (Metal / Vulkan / CPU, single file)

by infosave - opened about 16 hours ago

Discussion

infosave

about 16 hours ago

Hi XHToken team, thanks for Spark-X2.6!

I've added Spark-X2.6-4B support to cortiq, a Rust inference engine with no Python and no ML framework, and published the converted weights of Spark-X2.6-4B (the post-trained model built on this base) as single CMF files: https://huggingface.co/infosave/Spark-X2.6-4B-cmf

The whole architecture runs natively: 9 full-attention + 27 sliding-window (512) layers, exact-GELU FFN, tied head, 1M-token context, thinking mode and tool calling through the embedded chat template. GPUs: NVIDIA / AMD / Intel through Vulkan or DX12, Apple silicon through Metal; the CPU as well.

Files (quantized from the bf16 checkpoint)

file	size	wikitext-2 ppl
`spark-x2.6-4b-q8_2f.cmf` (8-bit, default)	4.14 GB	10.71
`spark-x2.6-4b-q4tp.cmf` (4-bit)	2.16 GB	11.88

ppl: first 4096 tokens of wikitext-2 test, BOS in front, cortiq ppl --tokens 4096.

Decode speed, one stream, tok/s

hardware	backend	q8_2f	q4tp
RTX 2000 Ada 16 GB	Vulkan	41.5	67.2
Mac mini M4 24 GB	Metal	20.1	34.8
Xeon E5-2699 v4, 6 cores	CPU	4.5	4.8

Prompt processing at 1,000 tokens: 370 / 351 tok/s on the RTX 2000 Ada, 286 / 244 tok/s on the M4 (q8_2f / q4tp).

Long context: the sliding layers keep only their window, so the KV state is ~72 KB per token (1.24 GiB at 16k tokens). cortiq serve exposes an OpenAI-compatible API, and each chat turn prefills only what changed: with a 2,300-token history the next turn starts in ~0.6 s on the M4.

Try it

cargo install cortiq-cli
hf download infosave/Spark-X2.6-4B-cmf spark-x2.6-4b-q8_2f.cmf --local-dir .
cortiq run spark-x2.6-4b-q8_2f.cmf --prompt "Explain quicksort in three sentences." --no-think

Prebuilt binaries for Linux, macOS and Windows: https://github.com/infosave2007/cmf/releases (cortiq 0.8.43 or later). The weights stay under your Apache-2.0 terms.

Feedback welcome!

Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.

Tap or paste here to upload images

· Sign up or log in to comment