Ollama v0.40.15 Pre-release: In-Process Loading of CUDA Libraries and MiMo V2.6 Model Support

A Minor Release Iteration Focused on Underlying Infrastructure

Ollama has released a pre-release version v0.40.15 (tagged v0.40.15-rc2). The commit history shows that this update contains only two functional changes: on Linux, the MLX backend now loads bundled CUDA libraries in-process, and support for the MiMo V2.6 model has been added. The number of changes is small, but the direction is clear—continuing to lower the barrier of environment dependencies for local inference while expanding the list of out-of-the-box models.

Judging from the version number and commit tags, this is still in the RC (Release Candidate) stage, a validation milestone before the official release. Users prioritizing stability may want to wait for the official version.

MLX and CUDA on Linux: Bringing Library Loading Into the Process

The most noteworthy technical change is the commit "mlx: load bundled CUDA libraries in-process on Linux" (#19433) by contributor dhiltgen.

Previously, when using MLX-related capabilities on Linux, CUDA library discovery and loading often depended on the system environment, making users prone to issues such as library path mismatches, version conflicts, or missing dynamic libraries. By loading the CUDA libraries distributed with the package in-process, the runtime no longer fully depends on the host machine's CUDA installation state, reducing configuration overhead.

The value of such a change lies not at the feature level but at the deployment experience level. For developers running local inference on Linux servers or development machines, "install and it works" is more appealing than an extra feature. It also makes the distribution package more self-contained, reducing irreproducibility issues caused by environment differences.

MLX is an open-source machine learning array framework introduced by Apple, originally designed for the Metal backend on Apple Silicon to efficiently perform model inference and training locally. It is not exclusively tied to macOS: the community and official teams have gradually expanded support for other platforms, including running on Linux where GPU compute is involved. CUDA is NVIDIA's GPU parallel computing platform and programming interface. If MLX needs to utilize NVIDIA GPUs on Linux, it must perform underlying computations through CUDA libraries (such as cuBLAS, cuDNN, NCCL, etc.).

"In-process loading" means the program directly loads and links these CUDA dynamic libraries distributed with the release package during its own runtime, rather than relying on the operating system's dynamic linker to search from system global paths (such as /usr/lib or locations specified by LD_LIBRARY_PATH). Multiple inference frameworks in the ecosystem, such as PyTorch and ONNX Runtime, have adopted or evolved toward similar strategies to make distribution packages self-contained and avoid runtime errors caused by inconsistent library versions across machines.

MiMo V2.6 Model Support

The second change is the addition of MiMo V2.6 model support (#19400), also submitted by dhiltgen. A model entering the official support list means users can directly pull and run it through Ollama without having to handle format conversion and inference adaptation themselves.

From the pace of updates, Ollama follows new models quite quickly, which has become one of its core competitive advantages as a local inference entry point: new progress on the model side can be quickly translated into a single command executable by end users.

Ollama's model support is not "any model file works out of the box." Models in the official support list typically mean: model weights are converted to Ollama's Modelfile and chunked formats such as GGUF, and the accompanying prompt templates, tokenizer behavior, context length, and quantization options are all verified to work, and can be directly pulled via ollama run model-name. Once a model enters the official list, users can avoid the entire process of downloading weights, handling format conversion, and debugging chat templates themselves.

It is worth noting that minor version iterations in model version numbers (such as from V2 to V2.6) often require synchronized updates to tokenizer configuration and quantization support in official support. The work is not just "adding a name"—support for minor version models also requires independent adaptation and validation commits. Quickly following new releases on the model side is key for local inference tools competing for developer entry points.

How to View the Pre-release Version

This version is marked as Pre-release, with the full change comparison at v0.40.14...v0.40.15-rc0. In terms of community feedback, the release page received a dozen or so likes and a few rocket emojis, indicating there is attention but not major-version-level buzz.

For production environments, it is recommended to wait until RC validation is complete and the official version lands before upgrading. For developers who want to validate CUDA library loading behavior or MiMo V2.6 performance in advance, they can try it in an isolated environment first, focusing on performance differences under multi-GPU, containerized, and different CUDA driver versions.

RC (Release Candidate) is a stage in the software release process: features are frozen, and remaining work focuses on fixing blocking issues and validating stability, with the official version ultimately selected from several RCs. rc2 means this is the second candidate version, indicating there were still issues to fix after rc1. For open-source projects, pre-release versions are usually not pushed to the stable update channel, and package managers or automatic upgrades may not pull them. Users need to explicitly specify the version or tag to install.

In addition, the change comparison range v0.40.14...v0.40.15-rc0 ends at rc0, meaning subsequent fixes in rc1 and rc2 may not be fully reflected in that comparison. When evaluating, it is best to refer to the commit records of the tags corresponding to each RC.

Summary

v0.40.15 is a typical "small steps, fast pace" update: one change fixes the environment, one adds a model. It does not change Ollama's product positioning, but continues to solidify the two most critical things for a local inference tool—less environment fiddling, more model support.

One-line Summary

Ollama v0.40.15-rc2 has only two changes: in-process CUDA library loading on Linux and added MiMo V2.6 support.

Paragraph Summary

Ollama released the pre-release version v0.40.15-rc2. The commit history shows only two functional changes: first, on Linux, the MLX backend now loads CUDA libraries distributed with the package in-process; second, support for the MiMo V2.6 model has been added. Both were submitted by contributor dhiltgen. The former addresses common Linux user issues such as CUDA library path mismatches, version conflicts, and missing dynamic libraries by making the distribution package more self-contained and reducing deployment and environment configuration costs; the latter brings the new model into the official support list, allowing users to pull and run it directly without handling format conversion and inference adaptation themselves. The article notes this is a typical "small steps, fast pace" iteration that does not change the product positioning but continues to solidify the two key capabilities of a local inference tool. Given that it is still in the RC stage, production environments are advised to wait for the official version, while developers can validate performance in advance in isolated environments under multi-GPU, containerized, and different CUDA driver versions.