Ollama v0.34.4 Speeds Up Structured Outputs and Enhances Apple Silicon Performance
What changed
Ollama has released version 0.34.4, delivering targeted speed improvements for Apple Silicon hardware, core execution engine updates, and key fixes for local model management.
A primary enhancement in this update addresses structured outputs on thinking models. Output constraints now apply in a single pass rather than through multiple evaluation steps, increasing overall generation speed and improving adherence to defined schemas.
For macOS users and Apple Silicon architectures, v0.34.4 introduces model-specific performance gains and visual processing upgrades:
- Prompt processing throughput is faster when running Qwen 3.8 models on Apple Silicon.
- Gemma 4 vision execution on Apple Silicon now dynamically selects the optimal image resolution on a per-image basis, retaining finer visual details for higher-resolution images.
- A bug that caused the macOS desktop application to hang or freeze while checking for active ChatGPT or Codex background processes has been fixed.
In addition, this release resolves an issue where setups containing large local model libraries would intermittently trigger false “model not found” errors during execution.
Under the hood, Ollama v0.34.4 updates three core upstream components: llama.cpp, Apple’s MLX runtime, and the XGrammar engine used for structured output generation.
Why it matters
Applying structured outputs in a single pass streamlines workflows that require reasoning models to output strict JSON schemas or constrained formats. Multi-pass enforcement often introduces latency overhead and potential formatting errors. Consolidating this process into one step alongside XGrammar updates leads to faster, more consistent structured responses for local automated agents.
The Apple Silicon optimizations address both text ingestion and multimodal vision tasks. Quicker prompt processing on Qwen 3.8 reduces time-to-first-token during long-context input evaluations. Meanwhile, per-image resolution scaling for Gemma 4 ensures that complex visual inputs—such as fine diagrams or high-resolution documents—do not lose critical detail to uniform downsampling.
For developers managing extensive local repositories, fixing false model resolution errors stabilizes automated CLI scripts and API pipelines. Updating underlying execution runtimes like llama.cpp and MLX ensures Ollama remains aligned with upstream performance optimizations and hardware acceleration fixes.
The practical audience is narrower than the headline suggests: structured-output users benefit across supported hardware, while the Qwen 3.8 and Gemma 4 improvements are specifically relevant to Apple Silicon workflows. The release notes report targeted changes rather than a universal speedup, so compare the same model, prompt, schema, and hardware before claiming a throughput gain.
What to do next
To update to Ollama v0.34.4, pull the latest build through your local package manager or restart the desktop client on macOS.
Developers using schema-constrained outputs with thinking models should re-run their pipelines to evaluate performance gains from single-pass processing. Teams utilizing Gemma 4 for vision applications on Apple Silicon should test high-resolution image payloads to confirm detail retention. Complete code modifications between release v0.34.3 and v0.34.4 are available on Ollama’s GitHub release page. When comparing local-model claims across runtimes, use the AI model announcement evaluation checklist and the Transformers GGUF guide.
Sources
Share Article
If this article helped you, please share it with others!