XHToken/Spark-X2.6

What it solves

Spark-X2.6-4B is a small language model built to run directly on devices like phones or laptops while handling real-world agent tasks — such as using tools, multi-step planning, and software engineering — without relying on cloud servers. It aims to give capable AI that fits on-device, balancing performance with limited memory and power.

How it works

It keeps the same architecture as its predecessor Spark-X2.5-4B but improves capabilities through better post-training. This boosts coding, agentic behavior, long-context understanding, multilingual skills, and reasoning. It offers three configurable thinking levels — none, low, high — so developers can trade off reasoning depth against latency and inference cost. The model runs on many inference engines: Transformers, SGLang, vLLM, MLX, llama.cpp, Ollama, LM Studio, and Unsloth.

Who it's for

Developers and teams building on-device AI applications, especially those that need agent-like behavior (tool use, planning) on resource-constrained hardware without sacrificing capability. It is also suitable for researchers who want to fine-tune the model using Llama-Factory.

Highlights

• State-of-the-art performance among similarly sized open-source models, sometimes matching or beating ~9B-parameter models.
• Supports a context length up to 1,048,576 tokens when served with SGLang (requires enough memory).
• Controllable thinking budget with three levels: none, low, high.
• Available on Hugging Face, ModelScope, Ollama, and other platforms, with OpenAI-compatible APIs for SGLang/vLLM.
• Apache 2.0 license.