The Pulse
Ollama 0.40.4 Brings Decision Models to CLI, Multimodal Embeddings
Ollama’s 0.40.4 release, published October 12, adds terminal commands for decision models and support for multimodal embeddings through llama-server. The update also improves model reuse and expands Apple Silicon support for Kev decision mo
AI.info Team · Written with AI ·
Ollama’s 0.40.4 release puts decision-model tasks in the command line and adds support for multimodal embeddings through llama-server. Published October 12, the update also adds a direct terminal example: ollama run nimble --ask "Is a refund being requested?" < ticket.txt. The release notes describe decision models as tools for yes-or-no questions, text classification and rubric-based scoring.
Decision models move into the terminal
The new command gives users a route to ask a decision model a question about text without first constructing a request to Ollama’s decision API. In the example, the command sends the contents of a ticket file to Nimble and asks whether it requests a refund. That keeps the task focused on a defined answer rather than asking a general-purpose chat model to compose a free-form response.
Decision models work through Ollama’s /v1/systemone endpoint, which the documentation says requires version 0.35.0 or later. That interface accepts a state and one or more typed questions, returning structured choices, probabilities or scores rather than ordinary generated prose. Ollama describes the uses as classifying text, answering yes-or-no questions and scoring text against a rubric; the 0.40.4 command brings that work to a direct shell workflow.
Multimodal embeddings reach llama-server
Version 0.40.4 adds support for multimodal embeddings through llama-server and fixes loading multimodal embedding models with small context windows. The notes do not identify a newly released embedding model or specify which modalities are available through this server support. Embeddings represent input as vectors used in tasks such as search and retrieval; multimodal systems can place different types of input in a shared representation space.
Ollama’s model library describes EmbeddingGemma 2, built by Google DeepMind, as a multimodal model that maps text, images, video and audio into one 768-dimensional vector space. Its library entry provides an Ollama pull command and examples using the /api/embed endpoint. The release’s llama-server support is a runtime change; the release notes do not say that 0.40.4 itself introduces EmbeddingGemma 2.
More changes target Apple Silicon and inference
The release also adds support for Kev decision models in Ollama’s MLX runner on Apple Silicon. Ollama says its inference routes now reuse verified manifests and already-loaded models, and that it updated llama.cpp with native decision-model support. Those changes sit alongside improvements to Gemma 4 prefill performance on NVIDIA GPUs using the MLX runner, and adjustments to how that runner handles long Laya states.
Other additions are aimed at developers using Ollama’s integrations. The ollama launch codex command now shows locally installed Ollama models in Codex’s model picker, where users can switch models with /model. Ollama also says ollama launch can use usage credits without requiring an upgrade to a paid plan.
A broader route for structured model outputs
The headline change is not a new decision-model family but a new way to invoke decision models from the terminal, paired with backend work for multimodal embeddings. Developers can now put a defined classification or scoring question into a shell command, while the server change extends support for embedding models that handle more than text. The release notes leave model-specific compatibility and performance figures unstated; their concrete example is the Nimble command for checking a ticket for a refund request.