Ollama v0.40.9 Released with Apple Silicon Phonon Speech Support

- 1. At a Glance
- 2. Overview
- 3. Key Changes
- 3.1. Support for Phonon Speech Generation on Apple Silicon
- 3.2. Introduction of the “d3" Decision-Making Model
- 3.3. Flint Model Support for Linux and Windows
- 3.4. Improved Error Notification During Streaming Output
- 3.5. Fix for Configuration Propagation in ollama launch
- 3.6. Preservation of Thinking Settings in Anthropic-Compatible Web Search
- 4. Supported Models and Hardware
- 5. How to Get It
- 6. Releases Since Our Last Article
- 7. Related Articles
- 8. What to Read Next
- 9. Sources
At a Glance
| Item | Value |
|---|---|
| Repository | ollama/ollama |
| Version | v0.40.9 |
| Published | 2026-10-22 |
| License | MIT |
| Source type | Primary source (the publisher itself) |
Values determined by this site’s code when the information was collected. Dates are JST.
Overview
Ollama, a tool developed in Go and provided under the MIT license, has released its latest version, “v0.40.9". Ollama is a tool for easily setting up and running inference on diverse models such as Kimi, GLM, MiniMax, DeepSeek, gpt-oss, Qwen, and Gemma in your local environment.
The change in this version that has the largest impact on users is the addition of support for speech generation (Speech) using Phonon in Apple Silicon-equipped environments. This enables text and chat model responses to be played back as audio simultaneously with generation, making it possible to build applications utilizing speech synthesis endpoints.
Key Changes
Support for Phonon Speech Generation on Apple Silicon
Speech generation using Phonon has been added for Apple Silicon environments. In addition to traditional text generation functions, it is now possible to generate and play back audio in real time while producing model responses or specified text.
Additionally, an OpenAI-compatible /v1/audio/speech endpoint has been added, allowing external applications to retrieve audio streaming. This affects users looking to build voice dialogue systems or text-to-speech applications on Apple Silicon-equipped Macs.
To use it, run the command as follows:
ollama run phonon "Hello from Ollama."
Introduction of the “d3" Decision-Making Model
“d3", a decision-making model family from the vLLM Semantic Router, is now supported in Ollama. It allows passing not only text but also structured state data and images as inputs, enabling classification processing, Yes/No answering, and probabilistic scoring of choices. The model is available in 2B, 4B, 9B, and 27B sizes.
This affects users who want to perform text and image intent classification and routing processing in a local environment.
You can run it by specifying a question as follows:
ollama run d3 "A customer says their card was charged twice." --ask "Is this a billing issue?"
Flint Model Support for Linux and Windows
The “Flint" model, which previously only ran on macOS, has been expanded to run on Linux and Windows operating systems.
Users who want to use the Flint model in Linux or Windows server and desktop environments can now start using it without platform restrictions.
Improved Error Notification During Streaming Output
Behavior has been fixed for streaming requests in OpenAI Responses and Anthropic Messages when failures occur midway through generation.
The issue where processing would previously terminate quietly after partial output has been resolved, and generation failures will now be correctly reported. This affects developers building systems using streaming APIs.
Fix for Configuration Propagation in ollama launch
When running the ollama launch command, a fix was made so that cloud model context lengths and capability information are accurately conveyed to coding tools.
This resolves issues where inappropriate context lengths were applied or compaction settings malfunctioned. This affects users integrating cloud models with development tools.
Preservation of Thinking Settings in Anthropic-Compatible Web Search
In the Anthropic-compatible web search feature, follow-up requests have been changed to use the thinking settings supported by the model.
This prevents avoidable processing failures that occurred across consecutive requests, improving the stability of dialogues utilizing web search.
Supported Models and Hardware
This version introduces support for new models and execution environments. Specifically, the following model architectures and hardware environments have been added to the supported list.
Phonon Speech Generation Model (Apple Silicon Environment)
The Phonon speech generation model is now available in Apple Silicon (Macs equipped with M-series chips) environments. It is now possible to generate and play back text and chat model responses as audio data, enabling local execution of speech synthesis. In addition, audio output can be utilized from various applications via the streaming-compatible, OpenAI-compatible /v1/audio/speech endpoint.
“d3" Decision-Making Model Family
Support has been added for various sizes of d3, the vLLM Semantic Router decision-making model family. By passing text, structured state data, or images to the model, users can locally execute tasks such as classification processing, Yes/No determinations, and probabilistic scoring for choices. The available model sizes are as follows:
- d3 2B
- d3 4B
- d3 9B
- d3 27B
Users can select the appropriate parameter size according to their use case and computational resources (such as VRAM capacity) to perform inference.
Flint Model (Linux and Windows Environments)
The Flint model, which previously ran exclusively on macOS, now also supports Linux and Windows OS environments. As a result, users running Ollama in Linux servers or Windows environments can set up and run the Flint model without OS restrictions.
How to Get It
For details on installers and update procedures in this release, please check the official release page.
For newly supported models, you can pull models and start inference using the following commands:
The execution command for speech generation using the Phonon model is as follows:
ollama run phonon "Hello from Ollama."
The execution command for decision-making by inputting text and questions using the d3 model is as follows:
ollama run d3 "A customer says their card was charged twice." --ask "Is this a billing issue?"
For detailed usage instructions and obtaining the latest binary, please refer to the official documentation and release pages.
Releases Since Our Last Article
Compiled by Local Model Watch from the project’s GitHub releases: the versions between this release and the last one we covered, which did not get separate articles. Full history: release tracker.
| Version | Released | Release notes |
|---|---|---|
| v0.40.8 | 2026-10-22 | GitHub |
| v0.40.7 | 2026-10-18 | GitHub |
| v0.40.6 | 2026-10-17 | GitHub |
Related Articles
- Ollama v0.40.12 Released: Adds Liquid AI LFM3 and d1 Models
- Ollama v0.40.15 Released with Parallel Vision GGUF Support
- Ollama v0.40.5 Released: NVIDIA Nemotron 3.5 Support Added
- Ollama v0.40.4 Released with Decision Models Support
What to Read Next
- Follow this tool → Ollama overview and release history (200 releases tracked)
- Other inference engines and runtimes → llama.cpp / vLLM / SGLang

