News Brief
Ollama 0.33.3: the changes to check before upgrading a local AI workflow
Ollama 0.33.3 adds cached-prompt token reporting and Gemma 4 image/audio support on MLX. For local-AI teams, those changes warrant an upgrade check—not a claim that every model runs faster.
Published · Updated · Sources reviewed 2026-09-04 · Event: 2026-09-02
For: Developers maintaining an existing Ollama deployment, especially MLX-based workflows.
What happened
The September 2 release lists cached-prompt token reporting, support for model-defined defaults in GGUF files, and Gemma 4 image and audio support on the MLX engine. It also updates underlying MLX, MLX-C and llama.cpp components. [release]
What to check in your own deployment
SourceVane analysis: These changes touch different parts of a deployment. Token reporting affects what an operator can observe; model defaults affect configuration; multimodal support concerns a specific engine path. Treating the list as one general performance improvement would erase those distinctions. [release]
For an upgrade trial, save the current version and explicit model options, then replay one representative request per workflow. Compare the returned content and settings before comparing speed. Keep the previous deployment available until those checks pass. This is a proposed test, not a test SourceVane has run.
What the announcement does not prove
The release notes are not an independent accuracy, latency or memory benchmark for your model and hardware. This brief does not establish a universal speedup or confirm that an existing production workload is safe to migrate.
Sources
- Ollama v0.33.3 release notes — What’s Changed
How this article was produced
SourceVane reviewed the linked documentation and separated reported facts from editorial interpretation. No hands-on benchmark was performed. External reader feedback has not yet been collected.
Revision history
2026-09-04 — Initial source-reviewed explanation. No hands-on benchmark performed.