Spark X2.5
An open-source lightweight large language model offered at a 4B parameter scale. It supports local deployment and is distributed by the community in quantized formats such as GGUF and MLX.
Timeline (last 90 days)
Spark X2.5 has reached 2,029,840 cumulative downloads, of which official channels contributed 1,019,681 and community derivative builds contributed 1,010,159.
Spark X2.5's official channels cover 6 platforms and 48 model repositories; on the community side, 109 projects independently publish derivative builds.
Official channels added a total of 167,689 downloads in the past 7 days, of which Hugging Face added 165,323 in the past 7 days.
The official source states in a footnote that all download figures are all-time cumulative downloads as of October 14, 2026, and that platforms do not define 'one download' uniformly. Download counts are not equal to the number of unique users.
The two Spark X2.5 models can be deeply adapted to Claude Code, Codex, and Open Cloud.
These two models were trained using entirely domestically produced computing power.
Spark X2.5 has only 4 billion parameters and a total download count of just over 3,000, but has received 467 likes.
All Facts (6)
Spark X2.5 has reached 2,029,840 cumulative downloads, of which official channels contributed 1,019,681 and community derivative builds contributed 1,010,159.
50%UnverifiedSpark X2.5's official channels cover 6 platforms and 48 model repositories; on the community side, 109 projects independently publish derivative builds.
50%UnverifiedOfficial channels added a total of 167,689 downloads in the past 7 days, of which Hugging Face added 165,323 in the past 7 days.
50%UnverifiedThe official source states in a footnote that all download figures are all-time cumulative downloads as of October 14, 2026, and that platforms do not define 'one download' uniformly. Download counts are not equal to the number of unique users.
50%UnverifiedThe two Spark X2.5 models can be deeply adapted to Claude Code, Codex, and Open Cloud.
50%UnverifiedThese two models were trained using entirely domestically produced computing power.
50%