OllamaFrom the builder
New model scheduling
Ollama introduced a much improved model scheduling system in which its new engine measures the exact memory a model needs before running it, instead of estimating as earlier versions did. Running gemma3:12b with 128k context on one RTX 4090, generation rose from 52.02 to 85.54 tokens per second, with all 49 layers on the GPU instead of 48. On two RTX 4090 cards with mistral-small3.2, prompt evaluation rose from 127.84 to 1380.24 tokens per second.
Ahead of running a model, Ollama’s new engine will now measure the exact amount of memory required compared to an estimation in previous versions of Ollama.
Saved . A shortened excerpt from the official blog; the original publisher retains authorship.