Google announced two new AI model releases at its annual I/O 2026 developer conference: Gemini Spark, a lightweight model optimized for on-device and edge deployment, and Gemini 3.5, an updated version of its flagship multimodal model featuring improved reasoning and coding performance.
Gemini Spark is designed for applications requiring low latency and offline capability, including mobile devices, embedded systems, and enterprise edge computing scenarios. Google stated that Spark runs fully on-device on Pixel 9 series hardware and select Android enterprise devices without requiring a cloud connection. The model supports text, image, and audio input in a 4-billion-parameter architecture that Google claims reduces inference cost by 78% compared to Gemini 2.0 Flash.
Gemini 3.5 targets enterprise workflows with an extended 2-million-token context window and updated performance across coding, mathematical reasoning, and long-document analysis tasks. Google reported that Gemini 3.5 scored 87.4% on the MMLU benchmark and achieved a 74.1% score on HumanEval coding tasks, improvements attributed to updated training data and reinforcement learning from human feedback.
Both models are available through Google Cloud Vertex AI, with Gemini Spark also accessible via the Gemini API. Google also announced updated enterprise pricing tiers and a new government-focused deployment track for Gemini 3.5 operating within FedRAMP High compliance boundaries.
Source: TechCrunch -- https://techcrunch.com/2026/05/19/google-gemini-spark-gemini-3-5-io-2026