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

Libraries

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

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

pipe = pipeline("text-generation", model="XHToken/Spark-X2.6-4B", 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", trust_remote_code=True, device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

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

Use Docker

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

SGLang

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

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

Spark-X2.6-4B and for the detailed answer on τ³-Banking in discussion #4. It made that number reproducible.

#21

by devzhr - opened 2 days ago

Discussion

devzhr

2 days ago

Hi SparkLLM team, thanks for releasing Spark-X2.6-4B and for the detailed answer on τ³-Banking in discussion #4. It made that number reproducible.

I'm trying to reproduce the SWE-bench results (Pro 54.3, Verified 62.8, Multilingual 64.7) locally with the official bf16 weights on vLLM + Spark-plugin, thinking=high, temperature=1.0, top_p=0.95, top_k=-1. Could you share the setup you used?

Agent scaffold and version (OpenHands, mini-swe-agent, SWE-agent, Claude Code, an in-house harness…)
Context window used during evaluation (128K, 256K, 1M?) and max output tokens per turn
Max steps / turns and any wall-clock timeout per task
Task set: the full public SWE-bench Pro set (731 tasks), Verified (500), Multilingual (300), or a subset?
Number of runs per task (pass@1 single run, or avg@k?)
Anti-hacking measures (network disabled, git history removed…)
For the starred Qwen3.5 numbers (SWE Verified 53.1 / 38.8, SWE Pro 33.8 / 29.4): which card or paper are they taken from? I couldn't find them on the Qwen3.5-9B / 4B model cards.
Thanks a lot!

pengwenzhi

SparkLLM org 2 days ago

•

edited 2 days ago

Hi @devzhr , thanks for your interest in reproducing our results! Below are the evaluation details for each benchmark.

SWE-bench Pro — 54.3

Agent scaffold: Claude Code v2.1.197
Context window: 1M
Maximum turns: 250
Evaluation metric: pass@1
Task count: Full 731 tasks Pro set
Per-task timeout: None configured
Anti-hacking measures: Git history removed; network blocklist covering GitHub-related domains.

SWE-bench Verified — 62.8

Agent scaffold: Claude Code v2.1.227
Context window: 256K
Maximum turns: 250
Evaluation metric: pass@1
Task count: Full 500-task Verified set
Anti-hacking measures: Git history removed; network allowlist; tool and package blocklists.

SWE-bench Multilingual — 64.7

Agent scaffold: Claude Code v2.1.227
Context window: 256K
Maximum turns: 250
Evaluation metric: pass@1
Task count: Full 300-task Multilingual set
Anti-hacking measures: Git history removed.

Source of the Qwen3.5 results

The starred Qwen3.5-9B / 4B results were taken from the Nanbeige4.2-3B report and are also listed in its model card.

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