Ollama v0.33.1
Ollama v0.33.1 is out. Claude on your Mac now talks to your local models, the MLX runner enforces JSON schemas, and large models load from slow disks without failing.
01 Claude, running on your models
Claude, running on your models.
Turn Claude on inside Ollama
Open Ollama, go to Apps, and switch Claude on. If Claude isn't installed on your Mac, Ollama downloads it and finishes the install for you.
You choose which model answers
In Ollama Settings, under Apps, pick an Ollama model for each Claude model option, then restart Claude. One model can cover more than one option.
What Claude keeps when it runs on Ollama
Claude keeps its own features
- Subagents — split a large task across agents
- Web search — runs through Ollama’s own web search
- Cowork — long coding and research jobs
- Auto mode — the agent decides when to ask
Claude keeps its own features here. Subagents split a big task up, web search comes through Ollama's own search, Cowork handles the long jobs, and auto mode decides when to ask you.
Switching it off puts Claude back
Turn Claude off in Apps and Ollama restores the configuration Claude had before. Quitting Ollama does the same thing. If Claude is open, Ollama asks to restart it.
The suggested pairings come from Ollama’s servers
The pairings Ollama suggests for each Claude model now come from Ollama's servers instead of a list baked into the app. So new recommendations reach you without an update.
The model picker holds its place
Picking a model in Settings used to jump the page. The picker was rebuilt on a popover, so the list opens where you left it and the scroll position stays put.
02 The MLX runner
The MLX runner.
A JSON schema is now enforced, not just accepted
A JSON schema is enforced, not just accepted
Ask the MLX runner for JSON and it used to hand back ordinary text. Now a grammar masks every sampling step, so each token it emits fits your schema.
What the constraint costs you
Constrained generation runs as fast as generating without a draft model. Speculative decoding is switched off while a schema is active, and that is the whole difference in speed.
Big models load off slow disks
Loading a large model off a slow volume used to abort with a command buffer failure. Metal was waiting on file reads, so the weights are now read on the CPU first.
03 Models and machines
Models and machines.
Qwen3.8 Flash Next runs on MLX
Qwen3.8 Flash Next now runs on Apple silicon through the MLX engine. Ollama added the architecture, its image tower, and the import step that shrinks its largest weights.
gemma3 stays where it can see images
The MLX port of gemma3 handled text only. Ollama removed it, so gemma3 keeps running on llama.cpp, where it can still look at your images.
Pi follows OLLAMA_HOST to a remote server
Ollama can launch the Pi coding agent. It kept whatever server address it had saved, so pointing OLLAMA_HOST at a remote machine did nothing. The launcher now rewrites it.
Under the hood
- llama.cpp — moved up to build b10630
- MLX — moved up a version too
- Linux Docker — the llama.cpp build is fixed
Both inference engines moved up a version, llama.cpp to build b10630. And the Linux Docker build, which broke on a missing file, works again.
Update to v0.33.1
That is Ollama v0.33.1. Update from ollama.com, or let the app update itself.
