SparkLLM releases Spark-X2.6-1.7B with 1M native context and day-zero vLLM support

What happened

SparkLLM’s Spark-X2.6-1.7B is out, a compact on-device agent model with 1M native context and controllable thinking (none/low/high). SparkLLM reports SWE Verified scores of 59.8 vs 62.8 for its 4B sibling and τ²-bench scores of 85.3 vs 86.9, both at high thinking. vLLM announces day-zero support via trust-remote-code, including vLLM-Ascend for NPU.

Read next

• More changes in today’s briefingGoogle DeepMind announces Genie 4 for real-time world simulation with native audio

2 reports

The story so far

SparkLLM released Spark-X2.6-1.7B, and vLLM announced day-zero support through trust-remote-code, including NPU support through vLLM-Ascend. vLLM describes it as a compact on-device agent model with a native context window of 1M tokens and three selectable thinking settings: none, low and high. SparkLLM reports that, at the high thinking setting, Spark-X2.6-1.7B scored 59.8 on SWE-Verified compared with 62.8 for its 4B sibling, and 85.3 on τ²-bench compared with 86.9 for the 4B model. Those reported results put the smaller model below its sibling on both benchmarks.