Spark-X2.6-4B

Spark-X2.6-4B is a 4.4B dense model built for on-device agents. It keeps the Spark-X2.5 architecture and gains its improvements from post-training: coding, tool use, multilingual tasks, and reasoning. Thinking is selectable per request through the chat template: none, low, or high.

These are Ollama-packaged GGUF quantizations of XHToken/Spark-X2.6-4B (Apache-2.0).

Files

Ollama

ollama run maternion/spark-x2.6:4b

ollama run maternion/spark-x2.6:4b --think=false

Thinking is on by default, so the model writes a reasoning block before the answer. Turn it off for a single request with "think": false:

curl -s http://localhost:11434/api/chat -d '{
  "model": "spark-x2.6:4b",
  "messages": [{"role": "user", "content": "What is the capital of Anhui Province?"}],
  "think": false,
  "stream": false
}'

With thinking enabled:

curl -s http://localhost:11434/api/chat -d '{
  "model": "spark-x2.6:4b",
  "messages": [{"role": "user", "content": "Plan a three-step migration from SQLite to Postgres."}],
  "stream": false
}'

Model details

• Parameters: 4.4B
• Architecture: spark2_5, 36 layers, hidden 2560, 16 attention heads / 4 KV heads, FFN 10240, vocab 131072
• Positional encoding: RoPE base 5,000,000 (sliding-window layers 10,000)
• Languages: English, Chinese
• Thinking modes: none, low, high
• Capabilities: completion, thinking, tools
• Sampling: temperature 1.0, top_p 0.95, top_k 0

Benchmarks

Full charts and comparisons: XHToken/Spark-X2.6-4B-GGUF benchmarks.

Links

• Base model: https://huggingface.co/XHToken/Spark-X2.6-4B
• GGUF files: https://huggingface.co/XHToken/Spark-X2.6-4B-GGUF
• SGLang: https://github.com/sgl-project/sglang
• vLLM: https://github.com/vllm-project/vllm
• MLX: https://github.com/XHToken/Spark-MLX-LLM
• Fine-tuning: https://github.com/XHToken/LlamaFactory

License

The Spark-X2.6 model series is licensed under the Apache 2.0 License.

Citation

If you find our work helpful, feel free to give us a cite.

@misc{sparkx2.6,
    title  = {Spark-X2.6-4B: Advancing the Frontier of On-Device Agents for Real-World Tasks},
    author = {SparkLLM Team},
    year   = {2026}
}