XHToken released Spark-X2.6-4B as its latest on-device language model. It keeps the Spark-X2.5-4B architecture but improves post-training. The model shows stronger coding, agent, long-context and multilingual performance. It matches or beats some 9B-parameter open-source models on benchmarks. It offers three thinking levels: none, low and high. It supports a 1,048,576-token context length. It runs on Transformers, SGLang, vLLM, MLX, llama.cpp, Ollama, LM Studio and Unsloth. It is licensed under Apache 2.0. Model weights and code are available on Hugging Face.
What changed
Spark-X2.5-4B used the same architecture but with lower post-training performance.
What it unlocks
Running stronger coding, agent, long-context and multilingual tasks on-device with configurable thinking levels.
- 4B parameters
- 1,048,576 token context length
- 85 downloads last month
What you need to act on it
- a device with sufficient memory for 4B parameters
- compatible inference framework (Transformers, SGLang, vLLM, MLX, llama.cpp, Ollama, LM Studio, or Unsloth)
Sources