Ollama shipped three releases in ten days this month. Version 0.34.0 landed on September 5, 2026, then 0.34.1 followed on September 14, and 0.34.2 on September 15. The pattern is familiar for the self-hosted command-line tool and local API for running open-weight language models: a headline feature first, then two rounds of cleanup that fix the rough edges. Between the three releases, you get a new bridge into ChatGPT Desktop, a fixed memory leak, and an API endpoint that’s suddenly a lot faster.
Ollama’s core pitch hasn’t changed: it wraps model downloads, quantization, and a REST-compatible API behind a single CLI. What the 0.34.x line does change is how close it now sits to closed-source clients like ChatGPT Desktop, plus a few operational annoyances that got fixed along the way. Here’s what’s in each release, working from 0.34.0 up to 0.34.2, the version currently shipping.
ChatGPT Desktop support headlines 0.34.0
The headline feature in 0.34.0 is support for using Ollama models directly inside OpenAI’s ChatGPT Desktop app. The release notes put it simply: the point is to let people “keep your existing workflow while running open models.” Instead of every ChatGPT Desktop conversation going to OpenAI’s cloud models, the same interface can now be pointed at models running locally through Ollama. Setup happens from the Ollama app itself, and for now that means macOS only.
It’s worth pausing on how unusual this is. A self-hosted, open-weight backend is now plugged into the official desktop client of one of the biggest closed-source AI vendors around, and the user doesn’t have to change how they work to use it.
0.34.0 also brought a few smaller changes: better structured output performance on Apple Silicon through MLX, support for OpenAI-compatible client-side tool search, and response compaction.
0.34.1: faster API, safer defaults
Nine days later, 0.34.1 showed up without a headline feature of its own. It’s the maintenance release that follows the big one. Here’s what changed:
- The MLX safetensors flow for
ollama createis no longer experimental. The trade-off: creating GGUF models now requires llama.cpp tooling for safetensor conversion and quantization. - MLX memory handling on Apple Silicon improved.
- The threshold for flagging runaway repeated-token loops moved up to 100 tokens, which cuts down on false positives. OCR workloads that legitimately repeat tokens were tripping this before.
- The
/api/tagsendpoint got a lot faster on large model libraries. In Ollama’s own testing, a cold start dropped from 3.1 seconds to 294 milliseconds. - The
typical_psampling parameter is deprecated. New models can no longer set it, though existing GGUF models still support it.
0.34.2: first-run setup and a memory-leak fix
Released the very next day, 0.34.2 is the version currently shipping. It adds two small features and fixes one real bug:
- Running the
ollamacommand for the first time now triggers a setup flow, asking whether to sign in or continue locally. That choice is now shared with the desktop app on both macOS and Windows. - A new
ollama://appsdeep link opens the desktop app’s Apps page directly. - Long generations using MLX speculative decoding were eating memory they shouldn’t have. That’s fixed now.
- llama.cpp got updated again.
Is it worth updating?
None of this is a security fix. It’s a feature and maintenance cycle, plain and simple. But there are still good reasons to move to 0.34.2 if you’re running Ollama:
- The ChatGPT Desktop integration is worth trying if you like that interface but want your requests handled locally instead of by OpenAI’s servers.
- The MLX speculative-decoding memory fix in 0.34.2 will matter a lot if you’re on Apple Silicon and run long generation sessions.
- The
/api/tagsspeedup from 0.34.1 is the kind of thing you won’t notice until you have dozens of models installed, at which point it’s hard to miss.
Updating follows the usual path: grab the latest installer from Ollama’s site, or pull the newest tag if you’re running it in Docker, then restart the service. None of the three release notes mention any migration steps.
Still, it’s worth skimming the full changelog before touching a production setup, especially the GGUF/safetensor tooling change in 0.34.1. If you’re new to the tool, our Ollama application page covers what it actually is, its license, and a one-line Docker install. Our earlier look at the 0.32.x release walks through the interactive agent CLI that shipped a couple of months before this. For the complete, authoritative changelog, Ollama’s official releases page on GitHub has all three: v0.34.0, v0.34.1, and v0.34.2.