RamaLama v0.26.0 Released: TLS Support and GPU Detection Updates

At a Glance
| Item | Value |
|---|---|
| Repository | containers/ramalama |
| Version | v0.26.0 |
| Published | 2026-11-04 |
| 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
Version v0.26.0 of containers/ramalama has been released. RamaLama is an open-source development tool written in Python under the MIT license that leverages container technology to simplify local AI model serving from various sources and production inference processing.
This v0.26.0 release introduces significant functional additions and improvements affecting user operations and security, such as added TLS communication support in llama.cpp and vLLM, extended API key settings for RAG pipelines, and optimized GPU detection processing.
Key Changes
Added TLS Support in llama.cpp and vLLM
TLS communication has been enabled in both the llama.cpp and vLLM backends. This allows for the encryption of communication paths for model inference servers. This affects users who want to serve local models while ensuring network security. Communication encryption settings can now be configured to match operational environments.
Extended API Key Management and Improved llama.cpp Router Mode
API keys for llama.cpp can now be output from the CLI, and the scope of the --api-key option has been extended to RAG pipelines. Additionally, a fix ensures that the --generate option is properly respected when running llama.cpp in router mode. This affects users building API environments or RAG pipelines that require authentication for inference. After updating, API keys specified during command execution are passed across all related pipelines.
Optimized GPU Detection and Expanded Information Display
The specification has been changed to automatically skip GPU detection processing when running subcommands that do not require it. Concurrently, the output of the ramalama info command now includes information on identified GPU and accelerator devices. Furthermore, an issue where GPU detection crashed on Windows environments without Podman installed has been resolved. Omitting unnecessary detection processing reduces command execution overhead and makes it easier to verify accessible hardware configurations.
RAG Feature and Networking Bug Fixes
Bugs related to port management and detach processing when using the --rag option, as well as an off-by-one error in port validation, have been fixed. Handling of model_alias for models specified in HTTP or HTTPS URL formats has also been improved. This affects users utilizing RAG features or operating by loading models directly from web URLs.
Backend llama.cpp Version Update
The internally used llama.cpp version has been bumped to v0.8.0 via v0.7.0. This affects all users running inference using llama.cpp as a backend. No special re-conversion or reconfiguration is required; users can update directly to use the new llama.cpp backend features.
Supported Models and Hardware
This version updates container environments and inference engines with hardware accelerator support.
First, the Docker image tag (quay.io/ascend/cann) for Huawei’s NPU (Neural Processing Unit) development toolkit, Ascend CANN (Compute Architecture for Neural Networks), has been updated to v9.2.0. This enables inference execution utilizing the latest CANN v9.2.0 runtime when running RamaLama containers on hardware environments equipped with Ascend accelerators. Users serving local models in Ascend hardware environments can build more stable inference environments thanks to the updated container base image.
In addition, the output of the system information command ramalama info has been expanded to include detailed information on detected GPUs and various accelerator devices in the report. Hardware devices recognized from the container side, such as NVIDIA CUDA-enabled GPUs, AMD HIP-enabled GPUs, Intel accelerators, and Ascend NPUs, are enumerated. This allows users to easily verify whether accelerators are properly recognized and passed through to the container in their execution environment.
Furthermore, with the internal inference engine llama.cpp updated to v0.8.0, it supports loading and running inference on various GGUF quantized model architectures supported by the latest llama.cpp. Environments operating existing GGUF models can execute inference using the updated engine without requiring special re-conversion or re-downloading of model files.
How to Get It
RamaLama can be updated using methods tailored to your operating system and deployment format. The main installation and update procedures are as follows:
Updating via pip (PyPI)
If you installed using Python’s package management tool pip, or in a Windows environment (using the WSL2 backend of Docker Desktop or Podman Desktop), use the pip command to update to the latest version.
pip install --upgrade ramalama
The standard command for a fresh installation is:
pip install ramalama
Updating on Fedora
If you installed from the Fedora package repository, use the dnf package manager to update.
sudo dnf upgrade ramalama
For a fresh installation, run the following command:
sudo dnf install ramalama
Updating via Installation Script (Linux / macOS)
If you are using the official installation script on Linux or macOS environments, you can fetch and install the latest version of RamaLama by running the following curl command:
curl -fsSL https://ramalama.ai/install.sh | bash
Updating via Standalone macOS Installer (.pkg)
If you use a self-contained installer containing Python and dependency packages on macOS, download the latest .pkg installer from the GitHub releases page. After downloading, double-click the file and follow the dialogs to install, or run the following command in the terminal to apply it to your system:
sudo installer -pkg RamaLama-*-macOS-Installer.pkg -target /
Related Articles
- RamaLama v0.25.0 Released: Security and Engine Updates
- NVIDIA Releases Dory1.0-4B: A Recurrent 4B LLM with Dynamic Loops
- MiniCPM-V-4.7-1B-DSpark-gguf Text Generation Model: 4GB+ VRAM
- MiniCPM-V-4.7-35B-A3B-DSpark-gguf Draft Model for Speedup: 4GB+ VRAM
What to Read Next
- Follow this tool → RamaLama overview and release history (68 releases tracked)
- Other inference engines and runtimes → llama.cpp / Ollama / vLLM

