Ollama 0.34.1: MLX Safetensors and the ChatGPT Desktop Bridge
Sources
Ollama v0.34.1 focuses on optimizing the Apple Silicon experience and expanding the "surface area" of local model consumption.
Technical Shift: MLX Safetensors
Moving MLX safetensors out of the "experimental" flag is a significant win for macOS users. By leveraging Apple's MLX framework natively, Ollama reduces the overhead of tensor conversion and improves memory handling. The shift to requiring llama.cpp tooling for safetensor conversion and quantization for GGUF models signals a maturation of the pipeline—moving away from "magic" internal scripts toward standardized community tooling.
The Wrapper Move: ChatGPT Desktop Integration
The ability to use Ollama models directly within the ChatGPT Desktop app is a strategic "wrapper" move. It allows users to keep a familiar UI while shifting the compute to local hardware. While this is great for accessibility, it hides the infrastructure. For the platform engineer, this means increased local resource contention (CPU/GPU) that is now triggered by an external application (ChatGPT) rather than a direct Ollama call, potentially complicating local resource quotas.
Verdict
The performance gains on Apple Silicon are real and should be utilized. The ChatGPT integration is a novelty that adds a layer of opacity to local resource usage—fine for individuals, but a potential headache for managed workstations.