Local LLMUpdates
Search
Release trackerBenchmark lab

Independent release intelligence for local AI

Local models change fast.The record should not.

A source-first record for people running language models on their own hardware. Scan verified release changes, then open the measurements when a result deserves a deeper page.

Open the release trackerInspect the benchmark lab

Release tracker

Verified changes, scoped to the exact release.

All updates
  1. 01

    macOS app and Claude Desktop proxy fixes

    Ollama restores following the system appearance, hands the macOS app off to an already-running instance, and prevents model-catalog updates from interrupting in-flight Claude Desktop proxy requests.

    verified
  2. 02

    Claude Desktop gateway, caching, and packaging fixes

    Ollama's v0.33.0 release lets Claude Desktop use Ollama as a third-party gateway provider, fixes prompt caching for recurrent-layer models, and fixes Linux and Windows packaging.

    verified
  3. 03

    Ollama reports v0.32.15 caching model metadata and normalizing Qwen 3.8 system messages

    Ollama reports that v0.32.15 caches resolved model metadata, fixes chat and generate wedging after a parser error, and normalizes Qwen 3.8 system messages.

    verified
  4. 04

    WebP transcoding and Qwen system-message handling

    The official notes list WebP transcoding for llama-server and Qwen renderer tolerance for non-leading system messages.

    verified
  5. 05

    Qwen3.8-27B support

    The release adds Qwen3.8-27B support, with Ollama stating particular optimizations for Apple Silicon.

    verified
  6. 06

    NVFP4 prefill and repeat-penalty defaults

    Ollama reports about 7–8% faster prefill on the Qwen3.6 and Muse Glimmer NVFP4 MLX models with a global scale, changes the default to 1.0 instead of 1.1 for models without `repeat_penalty`, and fixes a shared-digest blob-verification case.

    verified

Evidence standard

Useful only when the claim can be inspected.

Local AI changes by artifact, runtime, flags, and hardware. The publication treats those details as the story, not as footnotes.

  1. 01

    Source locked

    Material claims point to evidence a reader can open.

  2. 02

    Version scoped

    Results name the artifact, runtime, and environment they cover.

  3. 03

    Correctable

    Changes stay visible instead of silently replacing the record.

“A faster feed is easy. A trustworthy record of what actually changed is the product.”
Why this publication exists