Ollama: Launch Command Gets a Major Polish

Transcript

Hey there, developers! Welcome back to another episode of Ollama. I'm your host, and wow, do we have a fantastic update for you today. March 13th was absolutely buzzing with activity - 11 merged pull requests and some really thoughtful improvements that are going to make your daily workflow so much smoother.

Let's dive right into the star of today's show - the launch command is getting some serious love! ParthSareen has been on an absolute tear, and I'm genuinely excited about these changes because they solve some of those little friction points that can make or break your development flow.

First up, we've got a beautiful fix for headless integration with that `--yes` flag. You know how sometimes you just want to skip all the interactive bits and get straight to work? Well, now you can do `ollama launch <app> --yes` and it'll smartly use your previously set model without making you go through the selection dance again. It's one of those changes that feels small but saves you precious mental energy throughout the day.

But wait, there's more from Parth! The TUI model ordering was getting a bit wonky - it was incorrectly relying on checked models when it should have been static. Now when you uncheck a model from multi-select, it intelligently picks the closest model as the default. It's like having a thoughtful assistant who anticipates what you probably want next.

Moving into our cloud integration story - and this is where things get really interesting - BruceMacD tackled a gnarly issue with zstd-compressed requests hitting cloud models. Basically, when you were using something like the Codex CLI with cloud models, requests were just failing. The fix involved decompressing those zstd bodies before model extraction so everything flows smoothly to the cloud. It's the kind of behind-the-scenes work that just makes everything feel more reliable.

Bruce also cleaned up error handling in our middleware - no more cryptic JSON syntax errors when things go sideways. Instead of getting "invalid character 'e' looking for beginning of value," you'll get actual useful error messages. Small change, huge quality of life improvement.

Now here's a really cool technical fix from shivamtiwari3 - they solved this tricky Anthropic streaming issue where thinking blocks followed directly by tool use blocks were causing duplicate indices. The root cause was fascinating: the system wasn't properly closing thinking blocks when there was no text content in between. It's exactly the kind of edge case that can drive you crazy until someone sits down and really traces through the logic.

We've also got some nice documentation love - flipbit03 added proper docs for the reasoning_effort support in our OpenAI-compatible API. These are the thinking controls for models like qwen3.5, and now you'll know exactly how to disable thinking when you need to.

And speaking of cloud models, drifkin made sure that when you use the `/save` endpoint, it properly propagates the `:cloud` source so you don't accidentally try to run a local version instead. It's all about making the cloud integration feel seamless.

Oh, and Daniel added some important documentation for ROCm driver constraints - because nobody wants to spend their afternoon debugging GPU compatibility issues.

The testing coverage on all these changes is absolutely stellar. I'm seeing comprehensive test suites that cover edge cases and integration scenarios. This is how you build confidence in your releases.

Today's Focus: If you're using the launch command regularly, definitely update and try out that `--yes` flag for your automated workflows. And if you're working with cloud models, these compression and error handling fixes are going to make your life noticeably better.

The attention to detail in this release is just fantastic. From TUI polish to cloud integration robustness to proper error messages - it's all the stuff that makes software feel crafted rather than just functional.

Keep building amazing things, and I'll catch you in the next episode where we'll see what other improvements the team has in store. Until then, happy coding!