Spark X2.5-4B & 1.7B: the only on-device models with native 1M-token context — now open source
Spark X2.5-4B and Spark X2.5-1.7B, two on-device language models, have been released as open source. The announcement, posted on DEV Community, claims these are the only on-device models with native 1M-token context support. This means developers can now access models designed to run directly on devices while handling very long input sequences without external processing. The open-source release allows for integration into various applications, potentially reducing reliance on cloud-based inference. The models are positioned as a significant step for edge AI, offering a combination of on-device efficiency and extensive context window. The exact model names are Spark X2.5-4B and Spark X2.5-1.7B, indicating parameter counts of 4 billion and 1.7 billion respectively. The source does not provide additional technical details, benchmarks, or usage examples, so the digest focuses solely on the release and its stated unique feature.
Developers can now build on-device AI apps with native 1M-token context, a capability previously unavailable in open-source on-device models.