[Release] ollama/ollama v0.32.6

Here's the thing: improvements to Qwen3.5's performance on Apple GPUs sound promising, but what does 'faster' even mean without a concrete benchmark? The switch to using the model's MTP head for speculative decoding might enhance performance, but unless we see independent testing, it's hard to gauge the real impact. In the world of AI/ML, claims without evidence are just marketing fluff. Compared to OpenAI's offerings, you need to ask yourself if this performance boost justifies the shift, especially when you're already invested in tools that are proven to work at scale.

What they're not saying: while the new streaming format aligns with OpenAI's wire format, that doesn’t inherently make it better. It just means they’re catching up. If you’re already using OpenAI’s API or an established framework like Hugging Face, the incentive to switch isn't compelling unless you have specific use cases that directly benefit from Ollama’s architecture.

To be clear: the enhancements in this release may appeal to teams working heavily on macOS systems, particularly those eager to optimize resource usage on Apple hardware. However, those running multi-cloud or hybrid environments might find themselves better off sticking with more mature solutions. I’m not yet convinced that the hype for Qwen3.5 translates into real-world usability that outstrips established competitors.

The catch: while the update includes useful features like the correct reporting of truncated responses and new command options, these improvements feel incremental. If you’re eager to experiment with Ollama, it’s worth evaluating this version, but don’t expect it to revolutionize your workflow just yet. Keep your existing stack in mind before deciding to pivot based on this release alone.