Ollama v0.40.11 announces speech generation and the d3 decision models, but its own docs point to v0.40.9
AspireAI community · Ollama (GitHub release)
Ollama's v0.40.11 release notes, published October 24, announce two additions: speech generation on macOS through Gradium's Phonon text-to-speech checkpoints, streaming as it generates and exposed through the OpenAI-compatible /v1/audio/speech endpoint, and vLLM Semantic Router's d3 decision models in 2B, 4B, 9B and 27B sizes, where you hand the model text, structured state or an image and ask it to classify, answer yes or no, or score options with probabilities. One wrinkle is worth flagging: Ollama's speech documentation, last modified October 22, tells users to start on v0.40.9 or later, the Phonon library card says those packages require Apple Silicon and Ollama 0.40.9 or later, and third-party release trackers place both the Phonon speech section and the d3 models in the v0.40.9 line, while the v0.40.9 release body on GitHub itself lists only Flint on Linux and Windows plus streaming fixes. The honest reading is that these features exist and are documented as of October 24, but Ollama's release notes and its version history disagree about which release introduced them. Two other limits worth knowing: the default Phonon checkpoint is a 31M-parameter English model, speech models only generate speech so chat and embedding requests against them return a 400, and the d3 scores shown on Ollama's own library page are vendor-run rather than independently reproduced here.
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.