Ollama: Metal GPU Stability and Gemma4 Updates

Four pull requests were merged focusing on Metal GPU initialization hardening, Gemma4 renderer improvements, and VRAM-based model recommendations. The main highlight is a significant stability fix for Metal GPU systems experiencing kernel compilation failures.

Duration: PT1M36S

Episode overview

This episode is a short developer briefing from Ollama.

It explains recent repository work in plain language.

  • Show: Ollama
  • Published: 2026-04-30T00:00:00Z
  • Audio duration: PT1M36S

Transcript excerpt

This excerpt keeps the crawler page concise. Listen to the episode or use the RSS feed for the full update.

Good morning, this is your Ollama development briefing for April 30th, 2026.

Daniel Hiltgen merged a substantial Metal GPU stability fix addressing initialization failures. The change adds retry logic when Metal's tensor API probe passes but kernel coverage remains incomplete, which previously caused crashes during real kernel compilation. The fix also hardens error handling in Go's device…

Parth Sareen contributed two merged pull requests. The first updates Gemma4 renderers to support new templates from Google's latest models, including the 26B-A4B and 31B variants. The key improvement changes when the system determines if an assistant message has content, now checking after stripping Gemma thinking…

Bruce MacDonald cleaned up the codebase by removing an unused environment variable for Ollama update URLs that was only needed during development testing.

What's next: The Metal GPU hardening should improve stability for macOS users experiencing crashes, while the Gemma4 updates enable support for Google's newest model variants with improved content detection.

That's your Ollama update for today. Back tomorrow with more development news.

Nearby episodes from Ollama

  1. Go 1.26 Runtime Update
  2. Weekly Recap - MLX Threading & Model Recommendations
  3. MLX Threading Fixes and Claude App Integration
  4. Model Recommendations and Windows Gateway Fix
  5. Launch Experience Improvements and Model Recommendations
  6. Multi-Sequence Batching and New Model Support
  7. Tokenizer Bug Fix for BPE Processing
  8. Weekly Recap - MLX Performance & Launch Integrations