Ollama 0.40.9 adds speech generation, d3 decision models, and Flint on Linux/Windows

Ollama 0.40.9-rc3 adds three major features: speech generation via Phonon on Apple Silicon with OpenAI-compatible audio streaming, d3 decision models for classification and scoring tasks in 2B-27B sizes, and Flint cross-platform support extending to Linux and Windows. The release also fixes silent failures in streaming responses and improves cloud model context handling for coding tools.

What's new in Ollama 0.40.9

• Phonon speech generation: On Apple Silicon only; ollama run can play text or chat replies aloud during generation, with /v1/audio/speech endpoint for app integration
• d3 decision models: Ported from vLLM Semantic Router; supports text, structured state, or image inputs for classification, yes/no answers, or probability scoring; available in 2B, 4B, 9B, and 27B parameter sizes
• Flint expanded: Previously macOS-only, now runs on Linux and Windows
• Streaming fixes: OpenAI Responses and Anthropic Messages streaming now report generation failures instead of failing silently after partial output
• Cloud model context: ollama launch now correctly passes context lengths and capabilities to coding tools, fixing prior incorrect context and compaction settings
• Anthropic web search: Follow-up requests now respect the model's supported thinking setting

Why it matters

The d3 decision models bring structured reasoning capabilities previously requiring separate vLLM infrastructure into Ollama's local-first workflow, while cross-platform Flint support reduces deployment friction for developers building with Ollama on Linux servers and Windows workstations.