LoginSign Up
← Back to all sparks
O

Ollama

AI-ASSISTANTS
Velocity6.3

Run and manage large language models locally on your own machine.

ollama.com ↗

Ollama v0.40.11 adds local speech output on macOS, with an OpenAI-compatible audio endpoint.

local-llmtext-to-speechopen-modelsopenai-compatible-apispeculative-decoding
◆Current state
Ollama v0.40.11 can generate speech on macOS: ollama run reads text or a model's reply aloud as it is produced, and apps can stream audio from an OpenAI-compatible /v1/audio/speech endpoint using the Phonon voice model. The same release adds d3, a family of 2B–27B decision models that classify, answer yes or no, or score options with probabilities. v0.40.10 added NVIDIA's Nemotron 3.5 Bolt E4B with a speculative-decoding draft.
◆Where it's heading
Ollama is widening from a local text LLM runner into a local runtime for several model types: vision, decision scoring and now speech. The release candidates show the groundwork, with Phonon checkpoint support, a voice catalog and Gemma 4 draft heads landing over a few days before the stable cut.
◆Prediction
Expect speech output to reach Linux and Windows and more voices in the Phonon catalog; the decision-model work (PPLX Decider, d3) suggests more classifier-style models.

◆Recent moves

  1. 18h ago

    v0.40.13-rc0: LFM3 GGUF fixes and llama-server speech

    The first 0.40.13 candidate fixes LFM3 GGUF loading and wires speech through llama-server, with a llama.cpp bump. Follow-up fixes for the LFM3 and speech features that 0.40.11 and 0.40.12 shipped.

    View source ↗
  2. 1d ago

    v0.40.12: Liquid AI LFM3 models, 2B d1, faster MLX prompts

    The stable 0.40.12 ships what the two candidates staged: Liquid AI's LFM3 family (text, vision, audio and omni) and 2B d1 decision models, Apple-only for now, plus faster MLX prompts and DFlash2 drafts. It extends the speech release toward small local models for agents and tool calling.

    View source ↗
  3. 1d ago

    v0.40.12-rc1: Liquid AI LFM3 models and 2B d1 decision models

    The second 0.40.12 candidate adds Liquid AI's LFM3 family, including audio and omni variants, plus 2B d1 decision models, Apple-only for now. Small tool-calling and speech-capable models fit the local-agent direction set by the speech release, while the rest is MLX speed and OpenCode v2 support.

    View source ↗
  4. 2d ago

    v0.40.12-rc0: faster MLX prompts and DFlash2 drafts

    The first candidate after the speech release returns to inference speed: an MLX bump speeds up prompts for Gemma, Flint and Laguna, DFlash2 speculative decoding is added, and small-image detail is kept for compact Gemma 4 vision towers. It continues the speculative-decoding work started with Nemotron's draft in v0.40.10.

    View source ↗
  5. 3d ago

    v0.40.11: speech output on macOS and d3 decision models

    ⚡ SPARK

    The stable release that the Phonon release candidates were building toward: Ollama now speaks, not just writes. Together with the d3 decision models, it marks the shift from a chat-model runner to a general local inference runtime.

    View source ↗
  6. 5d ago

    v0.40.11-rc0: Phonon voice catalog

    Release-candidate plumbing that imports and lists Phonon catalog voices, plus docs. Groundwork for the speech feature that shipped in v0.40.11.

    View source ↗