270M nominal; 268,098,176 total dense parameters, including 167,772,160 tied embedding parameters
32K (32,768 tokens shared between input and generated output)
Gemma Terms of Use + Gemma Prohibited Use Policy (custom open-weight terms; not Apache 2.0)
Text-only, decoder-only Gemma 3 270M derivative fine-tuned specifically for function calling, released as google/functiongemma-270m-it in BF16 with Google-endorsed llama.cpp GGUF packaging. Uses 18 layers, width 640, grouped-query attention with four query heads and one KV head, and a 262,144-token vocabulary; fifteen 512-token sliding-attention layers alternate in five-to-one groups with three full-attention layers. Designed as a lightweight foundation for customized local agents, not a general-purpose direct dialogue model. Trained on public API definitions and tool-use interactions containing prompts, function calls, function responses and natural-language responses; the exact FunctionGemma adaptation objective, token budget and PT-versus-IT initialization checkpoint are not disclosed. Explicit training scope covers single-turn and parallel calls, not out-of-the-box multi-turn state tracking or dependent multi-step chaining; task-specific fine-tuning is recommended. Its custom developer/user/model chat format serializes tool declarations, calls and results using function-control tokens and <escape>-delimited strings rather than ordinary JSON output. The application must parse generated calls, execute tools and feed results back; neither the template nor tool training guarantees constrained decoding or automatic runtime execution. Google reports zero-shot BFCL Simple 61.6, Multiple 63.5, Parallel 39.0 and Parallel Multiple 29.5, with lower Live-subset results; these are individual subsets, not an overall BFCL score. On a held-out Mobile Actions evaluation, the released foundation scores 58%, rising to 85% after additional task-specific fine-tuning, not 85% for the unadapted checkpoint. A public completion-only supervised fine-tuning recipe supports customization; the separately adapted Tiny Garden LiteRT-LM demo illustrates offline game control. Google's mobile latency and memory figures concern dynamic-int8 specialized demos on a Samsung S25 Ultra, not the BF16 foundation. Supports local-first deployments and routing to larger models through an application scaffold. December 18 is verified availability-by through Google's explicit download offer and contemporaneous local-installation testimony; exact first public unlock remains unresolved.