Ollama v0.32.9 リリースOllama Releases v0.32.9

Ollama has published version 0.32.9 of its local model runtime, and the headline change is support for NVIDIA Nemotron 3.5 Lightning, a new open model aimed at agent workloads. For developers who prefer to run language models on their own hardware rather than through a hosted API, the update makes it easier to experiment with a mixture-of-experts design that is tuned for continuous, automated tasks.

NVIDIA Nemotron 3.5 Lightning is described as an open 30B mixture-of-experts (MoE) model with 3B active parameters. That split is central to why the model appears well suited to local execution. In an MoE architecture, the network is divided into many specialized subnetworks, or "experts," but only a small subset is activated for any given token. Here, the model holds 30 billion parameters in total while routing computation through roughly 3 billion at inference time. The practical effect is that the model can retain the breadth of a larger network while keeping the per-token compute and, to a degree, the runtime cost closer to that of a much smaller dense model. That profile is attractive for machines with limited memory and for workloads where responsiveness matters.

The excerpt positions the model at what it calls the execution layer of always-on agents. In agent systems, a model is not just answering a single prompt but repeatedly reasoning, calling tools, and acting in a loop that may run for extended periods. A model optimized for this layer is likely to prioritize consistent, low-latency responses over long, elaborate single answers. NVIDIA lists compatibility with agent harnesses such as OpenClaw and Hermes Agent, and the release notes indicate these are tied into NVIDIA's NeMo-based tooling. A harness, in this context, is the surrounding framework that manages an agent's prompts, memory, tool calls, and control flow, so support at that level is meant to make the model easier to drop into existing agent stacks.

For readers less familiar with the underlying platform, Ollama is a tool for downloading and running large language models locally, exposing them through a command-line interface and a local API that other applications can call. It packages models in a portable format and handles hardware acceleration where available, which is part of why new model support is often distributed as an incremental version bump rather than a separate product. The addition of Nemotron 3.5 Lightning fits that pattern: users can pull the model and run it through the same interface they already use for other open models.

The move also reflects a broader industry trend. Mixture-of-experts models have gained traction as a way to scale capability without a proportional increase in inference cost, with earlier examples including Mistral's Mixtral series and MoE variants from other labs. NVIDIA's Nemotron family, meanwhile, represents the company's push to offer open models alongside its hardware and software ecosystem, and pairing those models with agent-focused tooling suggests a strategy centered on always-on, tool-using systems rather than one-off chat. Making such a model available through Ollama extends its reach to the local and self-hosted community, where privacy, cost control, and offline operation are common motivations.

Some practical caveats are worth keeping in mind. A 30B total parameter count still implies meaningful storage and memory requirements even when only 3B parameters are active per token, so hardware needs will depend on the quantization level chosen and the specifics of the local setup. Performance claims around agent suitability come from the model's positioning rather than independent benchmarks, so real-world behavior across different harnesses and tasks will need to be verified by users. As with any newly supported model, capabilities, context length, and licensing terms should be checked against the official documentation before deployment.

In short, Ollama v0.32.9 is a targeted release that brings a compact, agent-oriented MoE model to local users. Those already building or testing automated agent pipelines are the most likely to benefit, and the update lowers the friction of trying an NVIDIA-backed open model without leaving a self-hosted environment.