Ollama: Fixing the Defaults That Silently Hurt You
Today's changes center on Ollama correcting defaults and detection logic that quietly degraded performance and reliability—wrong GPU memory readings on unified-memory AMD hardware, and a text penalty applied to models that never asked for it. Both fixes trace real, measured costs, not theoretical cleanup.
Duration: PT2M39S
Episode overview
This episode is a short developer briefing from Ollama.
It explains recent repository work in plain language.
- Show: Ollama
- Published: 2026-08-12T13:00:36Z
- Audio duration: PT2M39S
Transcript excerpt
This excerpt keeps the crawler page concise. Listen to the episode or use the RSS feed for the full update.
Good morning, it's August 12th, and here's your Ollama briefing.
The theme running through today's activity is Ollama tightening up assumptions that were quietly wrong for years, and once tested, turned out to cost real performance.
Start with PR 17679 from Jesse Gross. Ollama was applying a default repeat penalty of 1.1 to any model that didn't explicitly set one. Turns out almost nobody wants that. Llama dot cpp dropped the same default back in 2024, and libraries like vLLM and transformers don't apply one at all. The measured cost is the…
The second thread is hardware detection. PR 17685 from Speedway1 addresses AMD Strix Halo chips with unified memory, where the ROCm call for free GPU memory was actually returning system RAM instead of VRAM. That mismatch caused Ollama's scheduler to evict models it didn't need to evict, because it thought far less…
Related to hardware correctness, PR 17688 from Daniel Hiltgen re-enables the CUDA backend in CUDA builds for MLX, closing a regression left over from the image generation removal.
Elsewhere, model support keeps expanding: PR 17672 adds proper prompt layout handling for Nemotron 3.5 so it doesn't silently fall back to the 3.0 renderer,…
Nearby episodes from Ollama
- Silent Failures Get Louder
- Qwen Gets Serious About Coding Agents
- Launch Ecosystem Expands, Parser Edge Cases Get Squashed
- Backup Collisions and a Security Fix Converge
- Model Coverage and Tool-Call Reliability
- Vision Support Lands in the MLX Runner
- Weekly Recap - Vision Runners and Robustness Cleanup
- Edge Case Cleanup Across the Stack