XHToken has released Spark-X2.6-4B, a compact language model with 4 billion parameters designed to run directly on the device. It arrives a little over a month after Spark-X2.5 and keeps the same architecture: the capability gains come entirely from post-training, the refinement phase that follows base training.
The company claims substantial improvements in programming, tool use, multi-step planning and long-context handling, as well as in reasoning and in languages other than English. According to XHToken, the model achieves the best performance among open models of similar size and, on some benchmarks, matches or exceeds models of around 9 billion parameters.
Image: Hugging Face
Three reasoning levels
One of the most practical additions is an adjustable thinking budget. The model offers three levels, none, low and high, so an application can dynamically balance answer quality, latency and inference cost.
In its social announcement, SparkLLM reports some results at the high thinking level: 71.8 on BFCL-V4, 62.4 on MCP-Atlas, 53.1 on BrowseComp and 54.3 on SWE Pro. These are the company's own figures, to be read alongside the evaluation notes in the full table published with the release.
On efficiency, the claimed comparison is with on-device models such as Qwen3.5-4B and Gemma-E4B, with equal or better inference efficiency and stronger capabilities. The model is published on Hugging Face under an Apache 2.0 license and can be run with Transformers, vLLM, SGLang or Docker Model Runner. Quantized versions for llama.cpp, Ollama and LM Studio are also available.
The Spark family context
The release follows X2.5, presented as a series of compact models with a native context of up to one million tokens and support for more than 200 languages. Separately, iFlytek, which develops the Spark series, announced a next-generation model for its developer festival in Beijing on October 22, together with a cloud-device collaboration plan. Spark-X2.6-4B on Hugging Face and the iFlytek Spark X2.6 model presented in Beijing are two distinct announcements that the sources do not explicitly link.
Information about the model comes from the model card published by XHToken on Hugging Face, the GitHub repository and the post from the official SparkLLM account on X. License details and the cited results remain those stated by the authors.
No comments yet. Be the first!