Phi 3 (3.8B, 7B, 14B)

Generates coherent, instruction-aligned text responses using a decoder-only transformer architecture trained via supervised fine-tuning (SFT) and Direct Preference Optimization (DPO). Processes user messages in standard chat format (role/content structure) and produces contextually relevant outputs within a 4,096-token context window, optimized for latency-bound scenarios where model size and inference speed are critical constraints.

Extends the standard 4K context window to 128K tokens, enabling processing of long documents, extended conversation histories, and complex multi-document reasoning tasks. Accessed via specific model variant (phi3:medium-128k) requiring Ollama 0.1.39+, allowing developers to trade off some inference speed for dramatically increased context capacity without changing model weights or architecture.

Phi-3 models undergo Direct Preference Optimization (DPO) post-training to improve instruction adherence and incorporate safety measures, reducing harmful outputs and improving alignment with user intent. DPO uses preference pairs (preferred vs. dispreferred responses) to fine-tune the model without requiring explicit reward models, enabling instruction-following behavior that better matches user expectations while maintaining model efficiency.

Phi-3 training incorporates synthetic data generation to create high-quality reasoning examples (math, logic, code), enabling the small 3.8B model to achieve reasoning performance comparable to 7B-13B models trained on natural data alone. Synthetic data augmentation compensates for parameter count disadvantage by providing dense, reasoning-focused training examples rather than relying on scale.

Executes Phi-3 models entirely on local hardware (macOS, Windows, Linux, Docker) without sending data to external servers, using Ollama's runtime which handles model downloading, quantization format management, and GPU/CPU inference orchestration. Exposes both CLI interface (ollama run phi3) and HTTP REST API (localhost:11434) for programmatic access, enabling zero-latency, privacy-preserving inference with full control over model execution.

Deploys Phi-3 models to Ollama's managed cloud infrastructure (separate from local execution), enabling remote inference without maintaining local hardware while retaining API compatibility with local Ollama instances. Subscription tiers (Pro: $20/mo, Max: $100/mo) determine concurrent model capacity (1, 3, or 10 concurrent models), with identical REST API and SDK interfaces to local execution, allowing seamless switching between local and cloud deployment.

Phi-3 models are instruction-tuned and benchmarked on code generation, mathematical reasoning, and logical problem-solving tasks, leveraging synthetic training data and DPO post-training to improve reasoning capability. The 3.8B Mini variant achieves competitive performance on code and math benchmarks despite its small size, making it suitable for code completion, algorithm explanation, and structured problem-solving without requiring 7B+ parameter models.

Supports multi-turn conversations using standard chat message format (role: user/assistant, content: text), enabling stateless conversation management where each API call includes full conversation history. Ollama REST API and SDKs handle message serialization and streaming responses, allowing developers to build chatbot interfaces without managing conversation state or session persistence.

Generates text incrementally using HTTP Server-Sent Events (SSE), streaming tokens to the client as they are produced rather than waiting for complete generation. Reduces perceived latency and enables real-time UI updates (token-by-token display) without buffering entire responses, implemented via Ollama REST API with stream=true parameter.

Provides official Python and JavaScript SDKs that wrap Ollama REST API, enabling idiomatic language-specific code (async/await in JavaScript, context managers in Python) without manual HTTP request construction. SDKs handle message serialization, streaming response parsing, and error handling, reducing boilerplate and enabling integration into existing Python/JavaScript projects.

Phi-3 models can be deployed via Docker containers running Ollama, enabling reproducible, isolated execution environments across development, testing, and production. Docker images include Ollama runtime, model weights, and all dependencies, eliminating 'works on my machine' issues and enabling orchestration via Kubernetes, Docker Compose, or other container platforms.

Ollama enables selection between Phi-3 variants (3.8B Mini, 14B Medium) and context window options (4K default, 128K extended) via model tag syntax (e.g., phi3:latest, phi3:medium-128k). Developers specify desired variant in API calls or CLI commands, and Ollama automatically downloads and caches the appropriate model weights, enabling A/B testing or context-aware variant selection without manual model management.