Ollama 0.40.11 brings streaming local speech and d3 decision models

Ollama has published version 0.40.11 with two practical additions for developers running AI workloads outside a hosted chatbot. First, the release brings speech generation to macOS using Phonon, allowing the command-line interface to play generated audio as it arrives. Second, Ollama exposes d3 decision-model workflows for classification and scoring. Together, the updates expand the local-model runner from conversational text generation toward voice applications and small, narrowly scoped decision services.

The official release notes show a direct speech example using ollama run phonon, and an OpenAI-compatible /v1/audio/speech route through which applications can stream output. This is a macOS announcement, not a verified promise of identical speech support across Linux and Windows. The relevant distinction for builders is delivery mode: streaming can make spoken feedback feel more responsive, while actual startup delay and voice quality still need to be measured on the target device.

The new d3 pathway connects to vLLM Semantic Router's decision-model family. According to Ollama's release, it includes 2B, 4B, 9B and 27B sizes and can answer yes/no questions, classify a request or score choices using text, structured state or image inputs. A billing-triage example illustrates the scope well. A narrow classifier should be evaluated against carefully labeled production examples; a numerical score is not automatically a calibrated probability of real-world correctness.

These features are runtime integrations, not evidence that Ollama has created an entirely new frontier foundation model. Before deployment, pin the 0.40.11 version, test which d3 variant fits available memory, examine latency at concurrency, and check how the application handles failures and uncertain classifications. Teams using the audio API should also validate output formats, streaming behavior and operational permissions. The release is notable for making two additional local-model use cases accessible through familiar developer interfaces.