Phi 4 (14B)

Generates coherent, instruction-aligned text responses using a 14B-parameter transformer trained via supervised fine-tuning (SFT) on filtered synthetic and public domain datasets. The model processes English text input through a standard transformer decoder stack with 16K token context window, producing multi-turn conversational or task-specific outputs. Fine-tuning on curated instruction-response pairs ensures the model prioritizes explicit user directives over generic completions.

Executes multi-step reasoning tasks by leveraging transformer attention mechanisms trained on synthetic reasoning datasets and academic Q&A materials. The model decomposes complex logical problems into intermediate steps, maintaining coherence across the 16K token context. This capability is optimized through fine-tuning on reasoning-heavy datasets, enabling chain-of-thought style outputs without explicit prompting.

Processes input and generates output within a fixed 16,384-token context window using standard transformer attention mechanisms. The context window is a hard limit — inputs exceeding 16K tokens are truncated or rejected. Within this window, the model attends to all tokens with full attention, enabling coherent reasoning across the entire context but with quadratic memory complexity that limits window size.

Phi 4 is trained primarily on English-language data (synthetic datasets, public domain English websites, English academic materials) and optimized for English instruction-following and reasoning. The model has not been explicitly fine-tuned for other languages, though it may produce limited output in other languages due to exposure during pre-training. Performance degrades significantly on non-English inputs.

Executes model inference entirely on local hardware via Ollama runtime, streaming generated tokens in real-time to the client without round-trip latency to remote servers. The model is loaded into system memory once and reused across multiple inference requests, with streaming implemented via chunked HTTP responses or SDK callbacks. This architecture keeps all data local and enables sub-100ms time-to-first-token on typical consumer hardware.

Maintains conversation context across multiple turns by accepting message history in role/content format (user/assistant/system roles) and processing the full conversation history within the 16K token context window. The model uses standard transformer attention to weight recent messages more heavily than older ones, enabling coherent multi-turn dialogue without explicit state persistence. Conversation state is ephemeral — stored only in memory during the session.

Provides remote inference via Ollama Cloud, a managed service that hosts the Phi 4 model on Ollama's infrastructure with pay-as-you-go pricing. Requests are routed to geographically distributed servers (primarily US, with fallback to Europe and Singapore), and billing is based on tokens processed. Three pricing tiers offer different concurrency limits and usage quotas, enabling cost-scaling from hobby projects to production workloads.

Provides native SDK bindings for Python and JavaScript that abstract Ollama's REST API, enabling developers to integrate Phi 4 inference into applications without managing HTTP requests directly. The SDKs expose a unified `chat()` method that accepts message arrays and returns responses as objects or async iterables, with automatic serialization and error handling. Both SDKs support streaming responses via callbacks or async generators.

Enables inference via command-line interface (`ollama run phi4`) without requiring any SDK installation or programming. The CLI accepts prompts as arguments or stdin, streams responses to stdout, and supports interactive multi-turn conversations in a REPL-like interface. This capability is implemented as a thin wrapper around the local inference engine, making it suitable for shell scripts, automation, and quick prototyping.

Exposes Phi 4 inference via a REST API endpoint (`POST /api/chat`) that accepts JSON-formatted message arrays and returns responses as JSON objects. The API supports both streaming (chunked HTTP responses) and non-streaming modes, with standard HTTP status codes and error responses. This capability enables integration with any HTTP client library or tool (curl, Postman, etc.) without language-specific SDKs.

Phi 4 was trained using a blend of synthetic datasets (generated via automated processes), filtered public domain web data, and acquired academic materials, then fine-tuned with Direct Preference Optimization (DPO) to align outputs with human preferences. This training approach avoids reliance on large-scale human annotation while maintaining instruction-following quality. The synthetic data generation process is not publicly documented, but the resulting model exhibits strong performance on instruction-following and reasoning tasks.

Implements safety constraints and instruction adherence through Direct Preference Optimization (DPO) fine-tuning, which explicitly trains the model to prefer safe, instruction-aligned responses over unsafe or off-topic ones. The DPO stage uses preference pairs where safe/aligned responses are marked as preferred, enabling the model to learn safety constraints without explicit rule-based filtering. This approach is integrated into the model weights rather than applied as post-hoc filtering.