LoginSign Up
← Back to all sparks
O

Ollama

AI-ASSISTANTS
Velocity5.0

Run and manage large language models locally on your own machine.

ollama.com ↗

Ollama becomes a gateway provider for Claude Desktop — and this feed missed the release that says so.

local-llmgateway-providerclaude-integrationkv-cachemlxfeed-gap
◆Current state
This feed carries only release-candidate tags, and the last three shipped releases are absent from it: v0.33.0 (21 August), v0.33.1 (26 August) and v0.33.2 (27 August) are all live on GitHub and none was crawled. That matters because v0.33.0 is where the direction is stated — Claude Desktop can be configured to use Ollama as a third-party gateway provider — alongside a substantial prefill-caching correctness pass: cancelled prefills now keep every restore point they crossed, restore points no longer claim coverage they lack, and Claude Code's token-countdown system message was disabled because Ollama moved it to the front of the prompt and broke the KV cache on every request. The rc tags visible here carry only commit subjects, so what is stored is the shape of the work, not its substance.
◆Where it's heading
Ollama has moved from a runtime you call to the place a user's local and cloud models are organised for other clients to consume, and the gateway-provider framing in v0.33.0 is the clearest statement of it so far. The caching work points the same way: the bugs being fixed are the ones that only appear when a long-running agent client sits on top, cancelling and retrying prefills. Underneath, the MLX backend keeps gaining models and structured output support. Treat any read of this product from the feed alone as partial until the release entries are crawled.
◆Prediction
The three missing GA releases are the immediate gap, and until the crawler picks up the releases rather than the rc tags this feed will keep describing 0.33 as unreleased. On the product itself, the gateway work suggests further client configurations rather than depth on Claude Desktop alone.

◆Recent moves

  1. 5d ago

    App lists cloud account models for Claude

    The app lists a user's cloud account models for Claude. It is a candidate tag cut the day v0.33.2 shipped, and the release itself — dark mode restored, macOS handoff fixed, the Claude Desktop proxy no longer interrupting in-flight requests on a catalog update — is missing from this feed.

    View source ↗
  2. 6d ago

    macOS app handoff synchronization

    macOS app handoff synchronization, which lands in v0.33.2 as a fix for the app starting a second instance instead of handing off to a running one. Only the commit subject is stored here.

    View source ↗
  3. 7d ago

    Linux llama.cpp Docker build restored

    A Linux Docker build repaired after the build context missed a new cmake utility. Packaging correctness, and one of the few rc tags whose stored body says more than its title.

    View source ↗
  4. 7d ago

    MLX backend adds Qwen3.8 Flash Next

    Qwen3.8 Flash Next support in the MLX backend — another model on an existing engine. The v0.33.1 release it belongs to also added structured output to the MLX runner, which this feed does not show.

    View source ↗
  5. 7d ago

    Proxy preserves string content during image fallback

    The proxy preserves string content during image fallback. A narrow correctness fix in the request path the Claude Desktop integration depends on.

    View source ↗
  6. 8d ago

    Claude Desktop model mappings land in the app

    Claude Desktop model mappings arrive in the app — the candidate work behind v0.33.0, where Ollama is presented as a third-party gateway provider for Claude Desktop. The release stating that is not in this feed.

    View source ↗