SparkLLM has published Spark-X2.6-4B, a four-billion-parameter language model aimed at phones, laptops, and other local hardware, according to SparkLLM’s model card on Hugging Face. The weights are released under the Apache 2.0 license.
The card calls it an update to Spark-X2.5-4B, which arrived a little over a month earlier. The architecture is unchanged. SparkLLM credits better post-training for gains in coding, agent tasks, long context, multilingual use, and reasoning. Users can also choose one of three thinking levels (none, low, or high) to trade answer quality against speed and cost.
The headline claim is that the model is best in class among comparable sizes and matches or exceeds some 9B models on several benchmarks. The card’s text states no scores, and the comparisons are SparkLLM’s own, with no independent evaluation cited. It also says inference efficiency is on par with or better than Qwen3.5-4B and Gemma-E4B. The card credits “SparkLLM Team” as author and says nothing further about who is behind the lab.
With the weights free to download, the 9B-class claim is something a team can test against its own tool-calling workload in an afternoon.
Reported by Hugging Face (SparkLLM’s model card) on 21 October 2026.
