Ollama 0.34.1 makes MLX model import non-experimental
The update drops the experimental tag from MLX safetensors, routes GGUF creation to llama.cpp tooling, and cuts a cold /api/tags call to 294 ms.
In short
Ollama 0.34.1 takes the experimental label off creating models from MLX safetensors and hands GGUF conversion and quantization over to llama.cpp tooling.
At a glance
- MLX safetensors: ollama create is no longer marked experimental in version 0.34.1.
- GGUF: safetensor conversion and quantization now go through llama.cpp tooling.
- /api/tags: 3.1 seconds to 294 milliseconds cold in the project's own test, with consistent capability reporting.
- Repeat detection: a runaway loop is now called only after 100 repeated tokens, cutting OCR false positives.
- typical_p deprecated: unsettable on new models, still supported on existing GGUF models.
Ollama 0.34.1 takes the experimental label off creating models from MLX safetensors and hands GGUF conversion and quantization over to llama.cpp tooling. The release landed on September 14, 2026, following version 0.34.0.
Two formats, two directions
The release pulls MLX and GGUF apart rather than treating them as interchangeable inputs. On the MLX side, ollama create loses the experimental caveat it carried for safetensor weights. On the GGUF side, the runtime stops doing conversion work at all and points users at llama.cpp instead.
For anyone maintaining a build script, that is the change worth reading twice: a GGUF pipeline that relied on Ollama to convert and quantize safetensors now needs an llama.cpp step in front of it. Models already sitting in a local library keep working as they are.
The listing call got roughly ten times faster
The /api/tags endpoint enumerates locally installed models, which makes it the first request most front ends and scripts fire on startup. On a large model library, a cold call dropped from 3.1 seconds to 294 milliseconds in the project's own testing. Model capabilities are also reported consistently now, which matters to clients that switch UI features on the basis of that field.
Fewer false alarms on repeats
Runaway repeat detection now waits for 100 identical tokens before it intervenes. The stated reason is false positives in workloads such as OCR, where a model may legitimately emit the same sequence again and again. Raising the bar trades a slightly longer runaway loop for fewer wrongly truncated answers.
A sampling parameter on the way out
The typical_p sampling parameter is deprecated. It can no longer be set when a new model is created, while existing GGUF models keep support for it. The release also folds in updated MLX and llama.cpp builds.
What the notes do not say
The notes give no hardware and no library size for the 3.1-second baseline, so the speedup cannot be mapped onto an arbitrary local setup. The improved MLX memory handling on Apple Silicon comes with no figure at all. Nor do the notes say whether typical_p will eventually be removed outright.
FAQ
What changed for MLX safetensors in Ollama 0.34.1?
Creating a model from MLX safetensors is no longer flagged as experimental. The notes make no further stability commitment beyond dropping that flag.
Do I still need llama.cpp to build a GGUF model?
Yes. New GGUF model creation now expects safetensor conversion and quantization to be done with llama.cpp tooling before the model reaches Ollama.
Is typical_p still supported in Ollama?
Only for existing GGUF models. The parameter is deprecated and can no longer be set when creating a new model.
