XHToken publishes Spark-X2.6-4B on-device model weights
Original title: XHToken/Spark-X2.6-4B — new model trending #28 on Hugging Face
The 4B model keeps the Spark-X2.5 architecture and adds none, low, and high thinking levels.
XHToken released Spark-X2.6-4B weights and configuration in Hugging Face Transformers format, trending at number 28. The model keeps the Spark-X2.5-4B architecture and attributes capability gains to improved post-training for coding, agents, long context, multilingual use, and reasoning. The publisher says it matches or exceeds some roughly 9B models on several benchmarks and offers three thinking budgets: none, low, and high. Quickstart notes require transformers 4.57.1 and include SGLang serving examples for NVIDIA GPUs and Ascend NPUs.
Unconfirmed: Benchmark and efficiency claims are the publisher's own and are not independently verified here.
- Weights are for the post-trained model in Transformers format.
- Publisher compares inference with Qwen3.5-4B and Gemma-E4B.
- SGLang example sets a maximum context length of 1,048,576 tokens.
- Install command given is pip install transformers==4.57.1.
- Vendors
- XHToken
- AI models
- Spark-X2.6-4B
From the original article
6 min read at huggingface.cotext-generation · 85 downloads · 177 likes · license apache-2. 0
By XHToken · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.