Google: Gemma 3n 4B

Processes text, image, and audio inputs simultaneously through a unified transformer architecture optimized for mobile/edge deployment. Uses quantization and model compression techniques (likely INT8 or lower-bit precision) to reduce memory footprint while maintaining semantic understanding across modalities. Inference runs locally on device or via API without requiring cloud offloading for each request.

Implements a chat interface that follows user instructions and maintains conversation context across multiple turns. Uses a transformer decoder with attention mechanisms to track prior messages and respond coherently. The 'it' suffix indicates instruction-tuning via RLHF or supervised fine-tuning, enabling the model to follow complex directives, refuse unsafe requests, and adapt tone/style per user preference.

Generates text token-by-token using a quantized transformer decoder with optimized matrix multiplications for mobile hardware. Likely implements temperature scaling, top-k/top-p sampling, or beam search to control output diversity and quality. Inference is optimized for latency (sub-100ms per token on mobile) rather than throughput, enabling real-time interactive applications.

Exposes Gemma 3n via OpenRouter's REST API with HTTP POST endpoints for text generation and multimodal understanding. Requests are routed through OpenRouter's load balancer, which handles rate limiting, quota enforcement, and billing. Responses include usage metadata (prompt tokens, completion tokens, total cost) for cost tracking and optimization.

Gemma 3n applies post-training quantization (likely INT8 or INT4) and architectural pruning to reduce model size from ~12GB (full precision) to ~1-2GB (quantized), enabling deployment on devices with 4GB+ RAM. Quantization uses symmetric or asymmetric schemes with per-channel or per-token scaling to minimize accuracy loss. Inference kernels are optimized for ARM NEON (mobile) and x86 VNNI (laptop) instruction sets.

Generates responses that follow explicit user instructions (e.g., 'respond in JSON', 'use a formal tone', 'explain like I'm 5') by leveraging instruction-tuning via RLHF. The model learns to parse instruction tokens and adjust generation strategy accordingly. Attention mechanisms track both conversation history and instruction context to produce coherent, on-brand outputs.