Ollama is a name that keeps coming up while I’m tracking open-source AI trends. Anyone who’s run a local LLM has probably touched it at some point.

What stands out in the v0.33 line isn’t a new model. It’s that Claude Desktop now works as a third-party gateway. Running local models and a cloud app as two separate workflows always felt slightly annoying — if you’ve messed with local LLMs at all, you’ve probably hit that friction too.

This isn’t actually the first time Ollama tried this. Back in April, Anthropic quietly added a developer setting letting Claude Desktop connect to third-party inference gateways, and Ollama shipped ollama launch claude-desktop to take advantage of it almost immediately. It lasted days. A Claude Desktop update started rejecting non-Anthropic model IDs at the gateway level, and Ollama pulled the feature rather than ship something broken.

Here’s what’s actually in v0.33.1 (released August 26, 2026, with v0.33.2 already out as a release candidate), which fixes the mismatch properly this time with a dedicated local proxy instead of relying on the gateway setting:

  • Claude Desktop can now be added as a third-party gateway provider, mapping open-weight models like Qwen, DeepSeek, Kimi, or GLM into Claude’s Opus and Sonnet slots
  • Claude-specific model management landed in the app, set up entirely through Ollama’s own Mac app — select Claude, pick a model, done
  • MLX support for Qwen3.8 Flash Next on Apple Silicon
  • Structured output improvements
  • A GPU timeout bug fixed — it used to stall model loading on systems with slow storage

Two caveats worth knowing before you go looking for the toggle: it’s Mac-only for now, with no date given for Windows, and telemetry is off with zero data retention by default, which matters if part of the appeal for you is keeping things local in the first place.

Prefill handling also got more reliable under the hood. Cancel a generation mid-way and resume it, and it now picks up where it stopped instead of reprocessing everything from scratch.

Honestly, infrastructure fixes like that don’t hit you immediately. But the direction is pretty clear. Ollama’s been steadily turning ollama launch into a way to serve as the backend for other agent tools — Hermes Agent, GitHub Copilot CLI, and now Claude Desktop. This release fits that same pattern, and getting it right on the second try suggests they’re not treating it as a one-off stunt. Local models aren’t a separate track anymore; they’re just one more option you plug in.

Install is unchanged.

curl -fsSL https://ollama.com/install.sh | sh

Full release history is on the GitHub releases page.