Ollama v0.30.3 リリース — Gemma 4 12B モデルのサポートを追加 Ollama Releases v0.30.3

Ollama released version v0.30.3 on June 3, 2026, bringing a focused but meaningful update to its local LLM runtime: support for Google's Gemma 4 12B model. While the patch-level version bump might suggest a minor housekeeping release, the addition of gemma4-12b expands the roster of models available for local inference in a notable way.

The change was contributed by community member @pdevine via pull request #16457, reflecting the increasingly collaborative nature of Ollama's development. The gemma4-12b model belongs to Google DeepMind's Gemma 4 family, an open-weight series designed to run efficiently on consumer-grade hardware. The 12B parameter scale sits in a sweet spot that many developers and researchers favor — capable enough to handle complex reasoning tasks, yet manageable enough to run on systems with around 16GB of VRAM without requiring enterprise-class GPU infrastructure.

Ollama has become one of the go-to tools for running large language models locally, thanks to its straightforward CLI interface and cross-platform support across macOS, Linux, and Windows. Under the hood, it leverages llama.cpp as its inference backend, which enables everything from CPU-only inference to accelerated runs on NVIDIA and AMD GPUs. The `ollama run` command abstracts away model downloading, quantization selection, and runtime configuration, making it accessible even to users without deep ML engineering backgrounds.

The Gemma 4 series is understood to represent a generational improvement over Gemma 2, with refinements to both architecture and training methodology, though specifics vary by model size. Among the Gemma 4 lineup, the 12B variant offers a meaningful step up in capability compared to smaller 7B-class models, while stopping well short of the resource demands associated with 27B or larger configurations. This makes it a practical choice for local deployment scenarios where quality matters but hardware budgets are finite.

Ollama's model library already encompasses a broad range of popular open-weight models, including Meta's Llama series, Mistral, Microsoft's Phi family, and others. The steady pace of additions like this one reflects the vibrant open-weight model ecosystem that has taken shape over the past few years. For users prioritizing privacy, offline operation, or simply the ability to iterate quickly without API rate limits, each new model addition represents a concrete expansion of what's achievable on local hardware.

This release serves as a reminder that the local LLM space moves quickly, and tools like Ollama play a key role in making cutting-edge models accessible to a wide audience without cloud dependencies.