Ollama 0.40.11 teaches Phonon to read symbols, numbers and links aloud, imports its voices, and fixes two problems for coding tools

Ollama 0.40.11 is out, and most of what it changes is in how Phonon, its text-to-speech model, reads text aloud. Text that used to reach the model as raw bytes — symbols, email addresses, links — is now turned into words first , and Phonon can now carry a catalog of voices. Two fixes help coding tools: Qwen3-Coder tool calls that contain a literal closing tag no longer abort , and OpenCode started throughollama launchnow learns how much context the model really has. The installers were uploaded on 24 October 2026 at 17:27 UTC.

What changed

A note on the release notes.The page for 0.40.11 leads with two headings: speech — "Ollama can now generate speech on macOS" — and d3, which brings vLLM Semantic Router's decision models to Ollama. Neither is new in this release. Both notes were on the 0.40.9 page when we coveredOllama 0.40.9; that page now describes Flint on Linux and Windows , and the models' own library pages still say Phonon and d3 need Ollama 0.40.9 or later. So if you updated for 0.40.9, you already have both. What 0.40.11 adds is in the six changes merged between the two releases: 0.40.10 was made from commitf52b9e7, 0.40.11 from commit675150d, and the history between them runs through the changes below. The page's own changelog line still reads from v0.40.10 to v0.40.11-rc0, the pre-release this version grew from.

Phonon reads symbols, numbers and links as words.Phonon's tokenizer — the part that cuts text into pieces before the model sees it — has pieces for only 11 ASCII symbols and none for capital accented letters, so anything else reached the model as raw bytes. Ollama now ports the English text normalisation of Gradium's own runtime, xn-ptts (its 0.4.1 defaults), and fixes the readings that runtime gets wrong. In practice, symbols such asC#, snake_case, °C and Ctrl+C, as well as email addresses and links, are now read out, while Markdown markers, HTML tags and footnote marks are dropped. Times, dates, ratios and comparisons stay as written, as in Gradium's runtime, because the model reads them well. The change sits in the commit just before the release commit.

A voice catalog for Phonon. The release commit itself, 675150d, adds Gradium's voice catalog to Phonon. When you build a Phonon model withollama create, it imports any voices/NAME.safetensors files from the checkpoint folder, and ollama show lists them. Voice names work in any case, and asking for no voice means default. The model in Ollama's library is unchanged on this point: its default, Phonon 31M English, includes one voice, nameddefault.

Qwen3-Coder tool calls with a closing tag inside. When an argument of a Qwen3-Coder tool call contained the literal text </function>, XML decoding failed and the call was aborted. Ollama's qwen3coder parser now tracks which parameter it is inside, so a function closing tag inside a parameter value is treated as data. Calls that are genuinely broken — a missing or decorated closing tag, malformed nested edits — are still rejected.

OpenCode gets the real context length.OpenCode can compact too late when the model's available context is smaller than its advertised maximum. Compacting means summarising the conversation so far to make room; doing it late means the model runs out of room first. Now, when you start OpenCode withollama launch and pass it no arguments of your own, Ollama loads the model first, reads the context it actually allocated from /api/ps, and gives OpenCode that context and a matching output limit for that model. If you do pass OpenCode arguments, they go through unchanged and Ollama does not override OpenCode's choice of model. Your configuration files are not rewritten.

Under the hood.The bundled llama.cpp moves to build b11845 , and a change to the d3 code allows an empty state for vllm-d3.

What it means for you

If you use Phonon on an Apple Silicon Mac,update: replies that contain code names, units, addresses or links should now be read out as words. The normalisation is English , so it matters most for the English models. The library page's advice still holds: give the model text intended to be spoken. Phonon still needs Apple Silicon , so on Windows and Linux this release changes nothing for speech.

If you build your own Phonon models, the voice catalog is for you: put the voice files in a voices folder inside the checkpoint folder before ollama create, then run ollama showon the result to see which voices came in. If you only pulledphonon from the library, you still have the one defaultvoice.

If you run Qwen3-Coder as a coding agent, this fixes tool calls that failed whenever an argument contained </function>, for example in code that itself handles tool calls.

If you start OpenCode with ollama launch,update and launch it the usual way, without extra arguments, so Ollama can hand OpenCode the context the model really has.

How to update.The release's list of files carries an installer for each system, all uploaded on 24 October. On Windows, download and runOllamaSetup.exe; on a Mac,Ollama.dmg; on Linux, theinstall.shscript from the same list. Then open a terminal and typeollama -v: it prints the version you have installed, which should now be 0.40.11. If it shows an older number, quit Ollama completely, start it again and run it once more.

The source

The release page on GitHub, Ollama v0.40.11, its list of downloadable files with their upload times, and the v0.40.10 and v0.40.9 pages; the pull requests #19242 (voices), #19234 (Phonon text), #19228 (Qwen3-Coder) and #19193 (OpenCode); the commits 675150d, 547b98e, a5e91d7, f763dc7, 1472203 and 64a901e; and the library pages for phonon and d3.