Ollama: VRAM Fixes and Launch Polish
Proposed fixes would correct VRAM overcounting that evicts loadable models, while merged launch changes make ChatGPT restore and Pi and OpenCode setup safer. Additional fixes target Responses compaction, streaming errors, and Modelfile glob handling.
Duration: PT2M34S
Episode overview
This episode is a short developer briefing from Ollama.
It explains recent repository work in plain language.
- Show: Ollama
- Published: 2026-11-10T14:02:01Z
- Audio duration: PT2M34S
Transcript excerpt
This excerpt keeps the crawler page concise. Listen to the episode or use the RSS feed for the full update.
Hello, today is Tuesday, November 10th, 2026, and this is your Ollama developer briefing.
The big story is memory reliability: a proposed fix would stop the scheduler from overestimating video memory and evicting models that should fit.
First, GPU loading. PR 19632 tackles incorrect video memory accounting and GPU splitting, including a 4 to 8 times overcount of K V cache for hybrid recurrent models that caused unnecessary evictions, load failures, and unwanted splits across GPUs. Related, PR 19636 addresses Windows integrated GPUs, where keeping…
Second, the launch command. Merged PR 19583 makes ChatGPT restore ask before changing an active profile, leaves a closed app closed, preserves settings and history, and lets users complete the app's own quit prompt with cancel support. PR 19409 and PR 19634 extend that care to OpenCode and Pi, building thinking…
Finally, a cluster of correctness fixes. PR 19635 keeps compaction summary calls paired with their results to avoid a 400 on the next compaction in the Responses API. PR 19637 fixes the streaming client returning success on empty-body HTTP errors, and PR 19628 anchors relative Modelfile wildcards to the literal…
What's next:…
Nearby episodes from Ollama
- Tool Schema Fidelity, Desktop Launch, and Download Safety
- Tool Reliability and Mac Inference Tuning
- Faster Concurrent Inference and Unified Desktop Launch
- MLX Goes Parallel
- Stricter Model Imports and CLI Output Fixes
- Weekly Recap - MLX Model Expansion and Launch Polish
- Ripple on MLX and Validation Hardening
- MLX Memory Cuts and Tool-Call Fixes