Ollamav0.6

28 February 2025

Ollama 0.6 — Tool Use, Structured Outputs, 200+ Models

Ollama 0.6 brings function calling and structured JSON outputs to local models, making it viable for local-first agentic workflows. With 200+ models now in the library and Apple Silicon performance improvements, running production-grade inference locally is no longer a compromise.

• 1Tool use / function calling — compatible with OpenAI tool_calls format; works with llama3.3, qwen2.5-coder, mistral-nemo, and other supported models
• 2Structured outputs — constrain model responses to a JSON schema using /api/chat with format parameter
• 3OpenAI-compatible API extended — /v1/chat/completions now handles tools, format, and logprobs, enabling drop-in use with LangChain and LlamaIndex
• 4200+ models in the Ollama library including DeepSeek-R1, Qwen2.5-72B, Llama 3.3 70B, Phi-4, Gemma 2
• 5Apple Silicon performance — Metal optimisation reduces time-to-first-token by ~35% on M-series Macs for 7B+ models
• 6Parallel request support — multiple concurrent chat sessions without serialisation
• 7GGUF import improvements — import any GGUF model from HuggingFace or local file in one command
• 8Windows native GPU support (CUDA and DirectML) — no WSL required for GPU acceleration