Gemma 3 (2B, 9B, 27B)

Gemma 3 provides five parameter-efficient variants (270M to 27B) trained with Quantization-Aware Training (QAT), enabling 3x memory reduction compared to non-quantized models while maintaining near-BF16 quality. Models are distributed as GGUF artifacts via Ollama, supporting both local GPU inference and cloud-hosted deployment with automatic hardware optimization for NVIDIA Blackwell/Vera Rubin architectures.

Gemma 3's 4B, 12B, and 27B variants support multimodal input combining text and images, enabling visual question answering, image captioning, and document understanding. Images are encoded alongside text tokens within the transformer's 128K context window, allowing interleaved reasoning over both modalities without separate vision encoders.

Gemma 3 is claimed to have 'improved reasoning' compared to previous generations, implemented via standard transformer scaling (larger parameter counts, extended training) without documented architectural innovations. Reasoning improvements are claimed but not benchmarked; the mechanism is implicit in the model's training rather than explicit architectural features like chain-of-thought prompting or reasoning-specific loss functions.

Gemma 3 models are distributed as GGUF artifacts (Ollama's standard format), enabling efficient local storage and inference without requiring full-precision weights. GGUF is a binary format optimized for CPU and GPU inference; Ollama's runtime loads GGUF files and manages GPU memory allocation. Quantization-Aware Training (QAT) ensures quality parity with full-precision models while reducing disk and memory footprint by 3x.

Gemma 3's 4B, 12B, and 27B variants support 128K token context windows (32K for smaller variants), enabling multi-document reasoning, long-form summarization, and in-context learning with extensive examples. The extended context is implemented via standard transformer attention mechanisms without documented architectural modifications, allowing full document or conversation history to inform model outputs.

Gemma 3 is trained on data spanning 140+ languages, enabling text generation, summarization, and question-answering in non-English languages without language-specific fine-tuning. Language selection is implicit from input text; no explicit language parameter is required. Quality and coverage vary by language based on training data distribution, which is not publicly documented.

Gemma 3 models are served locally via Ollama's REST API (http://localhost:11434/api/chat), supporting chat completion format with streaming responses. The API abstracts model loading, GPU memory management, and inference scheduling, allowing developers to integrate Gemma 3 without direct CUDA/GPU programming. Requests are processed sequentially or in parallel depending on GPU memory availability and Ollama's internal scheduling.

Gemma 3 is accessible via Ollama's Python and JavaScript SDKs, providing language-native abstractions for chat completion, streaming, and model management. The SDKs wrap the REST API, handling serialization, streaming, and error handling. Python SDK supports async/await patterns; JavaScript SDK supports both Node.js and browser environments (via fetch).

Gemma 3 is available as cloud-hosted variants (gemma3:4b-cloud, gemma3:12b-cloud, gemma3:27b-cloud) via Ollama Cloud, with usage-based pricing tiers (Free: 1 concurrent model; Pro: $20/mo for 3 concurrent models; Max: $100/mo for 10 concurrent models). Requests are routed to Ollama-managed infrastructure; no local GPU required. Cloud models support the same REST API and SDK interfaces as local models, enabling seamless switching between local and cloud deployment.

Gemma 3 cloud models support tool calling via a schema-based function registry, enabling agents to invoke external functions (APIs, databases, tools) as part of reasoning chains. Tools are defined as JSON schemas; the model outputs structured function calls that are executed by the agent framework. This enables multi-step reasoning workflows where the model decides which tools to invoke and in what order.

Gemma 3 supports streaming responses via Ollama's REST API and SDKs, delivering model output in real-time chunks rather than waiting for full completion. Streaming is implemented via HTTP chunked transfer encoding; clients receive partial responses as they are generated, enabling low-latency user feedback and progressive rendering in UIs. Streaming can be disabled for batch processing or when full responses are required.

Gemma 3 models are managed via Ollama's command-line interface, supporting pull (download), run (execute), list (enumerate), and rm (delete) operations. Models are stored locally in a cache directory; pulling downloads the GGUF artifact from Ollama's registry. The CLI abstracts model versioning, GPU memory management, and process lifecycle, allowing developers to manage models without direct system administration.