Ollama v0.34.4 Released: Faster Structured Outputs and Stronger Apple Silicon Optimization

Ollama, one of the most popular tools for running large language models locally, continues its rapid release cadence. The latest v0.34.4 is a minor version update, but it delivers several practical improvements in inference performance, stability, and Apple Silicon compatibility. For developers and enthusiasts who rely on locally deployed LLMs, these optimizations have a direct impact on day-to-day experience.

Structured Output: Performance and Reliability Improvements

The most notable change in this update is the optimization of structured outputs for thinking models. Previously, generating structured outputs on reasoning models could require multiple passes, which hurt both speed and consistency.

v0.34.4 switches this process to a single pass, making structured outputs both faster and more reliable. For use cases where models need to return data in a fixed format — such as JSON for function calling, tool invocation, or data extraction — this improvement significantly reduces latency and improves output stability, cutting down on format parsing failures.

As more applications use local models as components in agent workflows, the quality of structured outputs directly determines the viability of downstream processes. These kinds of low-level optimizations may not be flashy, but they're critical for reliable production deployments.

Structured Outputs refer to forcing a language model to generate responses in a predefined format (most commonly a JSON Schema) rather than free-form text. This is typically achieved in one of two ways: dynamically filtering out invalid tokens during inference using grammar constraints, ensuring each generation step conforms to the target format; or guiding the model via prompt engineering and then parsing and validating the result. Thinking Models output an internal reasoning process before producing a final answer (usually wrapped in <think> tags). This "thinking" content doesn't need to conform to a structured format — only the final answer does. The previous multi-pass approach required handling the thinking and answer content in separate stages, introducing extra coordination overhead. The single-pass approach unifies both within the same inference loop, reducing intermediate state transitions and thus lowering latency while avoiding boundary errors that can arise when splicing formatted outputs.

Key Stability Fixes

This release also addresses two issues that affected the user experience:

• Intermittent "model not found" errors with large local model libraries: Users who store a large number of models locally occasionally encountered errors where a model couldn't be found. This has been fixed in the current version — a particularly important fix for power users managing many models.
• macOS app becoming unresponsive: The macOS client could freeze when detecting whether ChatGPT or Codex was running. This compatibility conflict has been resolved, improving desktop stability.

Both fixes address real pain points in everyday use, demonstrating that the Ollama team is actively polishing product stability alongside feature development.

Apple Silicon-Specific Optimizations

Apple Silicon (M-series chip) users represent a significant portion of Ollama's user base. This update brings two targeted improvements for that platform:

Faster Prompt Processing for Qwen Models

On Apple Silicon, prompt processing speed for Qwen 3.8 has been improved. Prompt processing is the phase where the model parses input before generating a response. Speeding up this stage means users get their first token faster, resulting in a smoother overall experience.

Better Image Handling for Gemma Multimodal

Gemma 4 on Apple Silicon now automatically selects the optimal resolution for each image, preserving more detail when processing high-resolution images. For multimodal tasks involving visual understanding, this translates to more accurate image recognition and better output quality.

These optimizations reflect Ollama's continued investment in the Apple ecosystem, making Macs with M-series chips an increasingly competitive platform for running local LLMs.

Core Dependency Updates

Beyond the above improvements, v0.34.4 also updates several key underlying dependencies, including llama.cpp, MLX, and XGrammar.

• llama.cpp is Ollama's core inference engine; updates to it typically bring performance improvements and support for new model architectures.
• MLX is Apple's machine learning framework, directly complementing the Apple Silicon optimizations mentioned above.
• XGrammar is closely related to structured outputs and grammar-constrained generation; its update reinforces the technical direction of the structured output improvements in this release.

Keeping these dependencies up to date ensures Ollama continues to benefit from the latest advances in the upstream open-source community.

llama.cpp is an open-source project initiated by Georgi Gerganov that implements efficient LLM inference in pure C/C++, enabling quantized models to run on CPU without a GPU. It supports the GGUF model format and has become a de facto standard in the local inference ecosystem — Ollama uses it as its core inference backend. MLX is Apple's machine learning framework designed specifically for Apple Silicon, taking full advantage of the M-series chips' Unified Memory architecture, which allows CPU and GPU to share tensor data with zero-copy overhead. This yields significant memory efficiency advantages in edge inference scenarios. XGrammar is a library focused on grammar-constrained generation. It compiles structured specifications like JSON Schema into efficient token masks, dynamically blocking invalid candidate tokens at each decoding step. This ensures output format compliance without sacrificing inference speed, making it a key low-level component for implementing structured outputs.

Summary

Overall, Ollama v0.34.4 is a release centered on refinement: no groundbreaking new features, but solid improvements across three dimensions — structured outputs, stability, and Apple Silicon performance. For users building applications on top of reasoning models, or running local LLMs on a Mac, upgrading to this version will deliver a noticeably faster and more stable experience. Users in these categories are encouraged to check the official release page and update at their earliest convenience.