Ollama 0.40.11 adds macOS speech synthesis and d3 decision models

Ollama version 0.40.11 introduces two major capabilities to its open-source local AI runtime: native speech synthesis on macOS and a new family of lightweight decision models called d3. The speech feature, invoked through the phonon model, streams generated audio in real time during ollama run sessions and exposes an OpenAI-compatible /v1/audio/speech endpoint for application integration. The d3 models, adapted from vLLM Semantic Router, enable classification, yes-or-no answering, and probabilistic scoring across text, structured state, and image inputs in 2B, 4B, 9B, and 27B parameter sizes.

Both features run entirely on local hardware, reinforcing Ollama's positioning as a self-hostable alternative to cloud AI APIs. The release also marks a first contribution from community member @SynapShift. Ollama remains under its existing open-source license with source code publicly available on GitHub.

What changes in Ollama 0.40.11

• macOS speech synthesis: The phonon model enables real-time text-to-speech generation on macOS, streaming audio as it is produced rather than waiting for full generation; this integrates with the existing ollama run CLI and the OpenAI-compatible /v1/audio/speech REST endpoint.
• d3 decision models: These models import vLLM Semantic Router's architecture for structured decision tasks, supporting classification, binary yes/no responses, and probability-weighted option scoring without requiring full conversational inference.
• Multimodal decision inputs: d3 accepts text, structured state data, or images as input contexts, expanding beyond traditional text-only classification pipelines.
• Model size variants: The d3 family ships in four parameter scales (2B, 4B, 9B, 27B) to match varying latency, accuracy, and hardware constraints on edge deployments.
• CLI integration: Both features are accessible through standard ollama run invocations with quoted prompts and optional --ask flags for decision queries, maintaining Ollama's established command-line workflow.

Why it matters

Local speech synthesis removes a common dependency on cloud TTS services for macOS-based AI applications, while structured decision models reduce inference costs for routing and classification workloads that previously required larger chat models. For developers building privacy-preserving or offline-capable AI pipelines on Apple Silicon, these additions narrow the feature gap with proprietary API offerings without requiring external network access or subscription accounts.