Ollama: Scheduler Memory Fixes and Faster Drafting
Three linked proposals tackle scheduler over-eviction from bad memory estimates on Metal and large-context models, while merged changes fix OpenCode v2 launches, tool-call edge cases, and add DFlash2 speculative drafting.
Duration: PT2M9S
Episode overview
This episode is a short developer briefing from Ollama.
It explains recent repository work in plain language.
- Show: Ollama
- Published: 2026-10-27T13:01:34Z
- Audio duration: PT2M9S
Transcript excerpt
This excerpt keeps the crawler page concise. Listen to the episode or use the RSS feed for the full update.
It's October 27th, 2026, and here's what's happening in Ollama.
The lead story is memory math. The scheduler has been kicking models out to free space the load never actually needed, especially on shared-memory Macs and with large-context models.
Three linked proposals from the same effort break that down. In pull request 19327, the predicted video memory estimate sizes every layer the same way, which badly misses models where layers differ — the example given is a 31-billion parameter Gemma at a 262-thousand context. Pull request 19324 finds a related…
The second theme is agent compatibility. Pull request 19310, now merged, fixes launching OpenCode version 2 by using standalone mode and tolerating its new config format, so the selected model actually sticks. Alongside that, pull request 19312 fixes tool-only turns from the Kimi command line client that send an…
Finally, performance. Merged pull request 19149 adds D-Flash 2 support for faster speculative drafting, proposing seven tokens per round while keeping output exact, alongside a routine llama dot cpp version bump.
What's next to watch is whether the memory fixes land and reduce evictions in testing, and how D-Flash 2 performs…
Nearby episodes from Ollama
- MLX Memory Safety and Silent Correctness Fixes
- MLX Performance Push and Compatibility Fixes
- OpenAI Fidelity and Runner Stability
- LFM3 Speech, Video and Reliability Fixes
- Mac Prefill Gains and Scheduler Fixes
- Weekly Recap - Phonon Voice Goes GA and MLX Speed Gains
- Gemma 4 Reliability and Runner Parity
- Tool Call Fixes and Phonon Voice Update