Ollama v0.40.4 Released with Decision Models Support

At a Glance
| Item | Value |
|---|---|
| Repository | ollama/ollama |
| Version | v0.40.4 |
| Published | 2026-10-13 |
| 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 v0.40.4 has been released. Ollama is a tool designed to easily run various open-weight models such as DeepSeek, Gemma, Qwen, Kimi, and GLM on local PCs and server environments. It is developed in Go and open-sourced under the MIT license.
The most significant change in this update is the official support for “decision models" in the command-line interface (CLI). This allows tasks such as text classification and yes/no judgments to be executed efficiently and directly from the terminal without having to construct complex prompts. It is a highly convenient update for engineers looking to incorporate LLM decision-making capabilities into automation scripts or existing workflows.
Key Changes
Native Support for Decision Models in the CLI
It is now possible to directly operate decision models through the ollama run command. This feature is specialized for tasks where specific questions are answered with “yes" or “no", text is classified into designated labels, or scoring is performed against specific evaluation criteria (rubrics). Specifically, it can be executed in a format like ollama run nimble --ask "Is a refund being requested?" < ticket.txt. This allows users to instantly obtain judgment results based on the contents of an input file. As the foundation for this feature, the internal engine llama.cpp has been updated to integrate native support for decision models.
MLX Runner Enhancements and Performance Improvements
Major improvements have been made to the MLX runner for Apple Silicon and NVIDIA GPU users. On Apple Silicon environments, the Kev decision models are now supported. Additionally, in environments using NVIDIA GPUs, the prefill (processing before inference starts) performance for Gemma 4 models has been improved. Furthermore, the reliability of speculative decoding has been enhanced. This includes proper handling of recurrent states after inference rollbacks occur, enabling more stable and faster inference. Also, instead of rejecting long Laya states, it has been changed to clip them and continue processing.
Support for Multimodal Embeddings
Multimodal embeddings are now supported via llama-server. This enables vectorization of data combining multiple modalities, such as text and images. In connection with this, a bug that occurred when loading multimodal embedding models with extremely small context windows has also been fixed, allowing a broader range of models to be used stably.
Streamlining and Speeding Up the Inference Process
The model loading and execution processes have been revised to improve inference performance. A mechanism to efficiently reuse verified manifests and models already loaded into memory has been introduced. This reduces the overhead when sending repeated requests to the same model or starting models multiple times in a short period, shortening the time to first token.
Enhanced Integration with Ollama Launch and Codex
The ollama launch command now allows the use of usage credits without upgrading to a paid plan. Additionally, as a developer-oriented feature, Ollama models installed locally are now displayed in the model selection screen of ollama launch codex. By using the /model command while running the Codex CLI, users can instantly switch between local models to compare and evaluate code generation behaviors.
Supported Models and Hardware
This release focuses more on expanding support ranges across existing hardware and runners rather than adding new model architectures.
The MLX runner running on Apple Silicon environments newly supports Kev decision models. This allows users on Macs with Apple Silicon to natively execute yes/no judgments and text classification using decision models.
For NVIDIA GPU environments, it is reported that prefill performance for Gemma 4 via the MLX runner has been improved. Users operating Gemma 4 on NVIDIA GPUs may experience shorter wait times before prompt processing begins.
Furthermore, with multimodal embedding models now supported via llama-server, models capable of generating embedding vectors combining images and text can now be handled by Ollama alone without requiring special external tools. Loading bugs for multimodal embedding models with small context windows have also been fixed, potentially allowing models that previously failed to load to operate.
All of these are internal runner enhancements, and the documentation does not state that users need to reconvert or re-download model files.
How to Get It
Specific installation commands inherent to this release are not listed in the documentation. Normally, Ollama is updated to the latest version via distributed installers or package managers, but since reliable procedures for this release cannot be confirmed from the documentation, please check the release page below for the latest procedures and detailed change contents.
ollama/ollama v0.40.4 Release Notes
The diff from v0.40.3 can also be checked on the Full Changelog page below.
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 |
|---|---|---|
| v0.40.3 | 2026-10-11 | GitHub |
Related Articles
- Ollama v0.40.12 Released: Adds Liquid AI LFM3 and d1 Models
- Ollama v0.40.2 Released: Automatic Model Upgrades & More
- Ollama v0.40.15 Released with Parallel Vision GGUF Support
- Ollama v0.40.9 Released with Apple Silicon Phonon Speech 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

