Spark-X2.6-1.7B
A compact on-device agent model introduced in the SparkLLM series. It has 1.7B parameters and natively supports a 1M context window along with a controllable thinking mode.
Timeline (last 90 days)
According to the official Twitter account @SparkLLM, Spark-X2.6-1.7B performs almost on par with its 4B-parameter version on the SWE-Verified and τ²-bench benchmarks.
Spark-X2.6-1.7B natively supports a 1M (million-level) context length.
Spark-X2.6-1.7B has controllable thinking capabilities, allowing users to switch among three thinking modes: none, low, and high.
The official information currently provides only the qualitative description "almost matches its 4B sibling" and has not yet published detailed benchmark scores or key metrics such as memory usage and latency during actual on-device runtime.
Spark-X2.6-1.7B gained vLLM support on its first day of release and can be loaded via trust-remote-code.
Spark-X2.6-1.7B also supports vLLM-Ascend and can be deployed on NPU-based hardware platforms.
Spark-X2.6-4B and Spark-X2.6-1.7B will soon be open-sourced, and the specific upgrades will be announced at release.
All Facts (6)
Spark-X2.6-1.7B natively supports a 1M (million-level) context length.
50%UnverifiedSpark-X2.6-1.7B has controllable thinking capabilities, allowing users to switch among three thinking modes: none, low, and high.
50%UnverifiedAccording to the official Twitter account @SparkLLM, Spark-X2.6-1.7B performs almost on par with its 4B-parameter version on the SWE-Verified and τ²-bench benchmarks.
50%UnverifiedSpark-X2.6-1.7B gained vLLM support on its first day of release and can be loaded via trust-remote-code.
50%UnverifiedSpark-X2.6-1.7B also supports vLLM-Ascend and can be deployed on NPU-based hardware platforms.
50%UnverifiedThe official information currently provides only the qualitative description "almost matches its 4B sibling" and has not yet published detailed benchmark scores or key metrics such as memory usage and latency during actual on-device runtime.
50%