--- title: I moved my local models from Ollama to llama.cpp, and... description: A long-time Ollama user explains why switching to llama.cpp improved their local AI workflow, despite similar raw token-per-second speeds. Key advantages cited... canonical: https://daily.dev/posts/i-moved-my-local-models-from-ollama-to-llama-cpp-and-the-reason-i-stayed-has-nothing-to-do-with-spe-qzdhhvk87 twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: I moved my local models from Ollama to llama.cpp, and the reason I stayed has nothing to do with speed | daily.dev og:description: A long-time Ollama user explains why switching to llama.cpp improved their local AI workflow, despite similar raw token-per-second speeds. Key advantages cited... og:url: https://daily.dev/posts/i-moved-my-local-models-from-ollama-to-llama-cpp-and-the-reason-i-stayed-has-nothing-to-do-with-spe-qzdhhvk87 og:image: https://api.daily.dev/og/posts/qzdhhvk87.png og:image:alt: I moved my local models from Ollama to llama.cpp, and the reason I stayed has nothing to do with speed og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # I moved my local models from Ollama to llama.cpp, and the reason I stayed has nothing to do with speed **[XDA Developers](https://daily.dev/sources/xda-developers)** · 4 min read · 1 upvotes · 0 comments ## Summary A long-time Ollama user explains why switching to llama.cpp improved their local AI workflow, despite similar raw token-per-second speeds. Key advantages cited include the --n-cpu-moe flag for fine-grained control over which model layers sit in system RAM versus VRAM (boosting throughput from ~75 to ~81 tokens/sec), the llama-bench tool for reliable, repeatable benchmarking, and more flexible model file management without duplicate downloads or import errors. The author argues Ollama remains better for beginners wanting a simple download-and-chat experience, but power users tinkering with models from Hugging Face will likely prefer llama.cpp's control and transparency. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### What does the --n-cpu-moe flag do in llama.cpp? It lets you choose how many layers of a mixture-of-experts model keep their experts in system RAM instead of GPU VRAM, giving finer control over memory placement than Ollama offers. In one test, offloading eight layers' experts to RAM increased throughput from 74.75 to 81.43 tokens per second on the same GPU and model, while Ollama has no equivalent setting. _Developers tuning local model performance can track llama.cpp configuration tricks like this via daily.dev._ ### Why do my local LLM benchmark numbers vary so much between runs? A large swing between runs, such as a 37% difference, is often caused by the first run still loading model weights into VRAM rather than measuring steady-state performance. Llama.cpp's built-in llama-bench tool addresses this by measuring prompt processing and generation separately, repeating each test five times, and reporting the average. _Anyone benchmarking local models reliably can follow discussions like this on daily.dev._ ### Does Ollama support offloading specific mixture-of-experts layers to RAM like llama.cpp does? No, Ollama currently has no equivalent to llama.cpp's per-layer expert offloading feature. A feature request for this capability has been open on Ollama's GitHub issue tracker since August 2025 without being implemented. _Developers comparing Ollama and llama.cpp feature gaps can follow updates on daily.dev._ ## Similar posts on daily.dev - [Ollama is the easiest way to start with local AI, and it's also the first thing I'd replace](https://daily.dev/posts/ollama-is-the-easiest-way-to-start-with-local-ai-and-it-s-also-the-first-thing-i-d-replace-vr6ueel1l) · XDA Developers · 1 upvotes · 0 comments - [Don’t Use Ollama for Local LLMs](https://daily.dev/posts/don-t-use-ollama-for-local-llms-szl4snpls) · Medium · 0 upvotes · 0 comments - [Ollama vs. LM Studio vs. llama.cpp: Which Local AI Runtime Should You Use in 2026?](https://daily.dev/posts/ollama-vs-lm-studio-vs-llama-cpp-which-local-ai-runtime-should-you-use-in-2026--vrytyob1e) · Machine Learning Mastery · 0 upvotes · 0 comments - [Friends Don't Let Friends Use Ollama](https://daily.dev/posts/friends-don-t-let-friends-use-ollama-tujmth8n1) · Hacker News · 3 upvotes · 0 comments --- Tags: [#self-hosting](https://daily.dev/tags/self-hosting), [#local-ai](https://daily.dev/tags/local-ai), [#ollama](https://daily.dev/tags/ollama), [#llama-cpp](https://daily.dev/tags/llama-cpp) [View this post on daily.dev](https://daily.dev/posts/i-moved-my-local-models-from-ollama-to-llama-cpp-and-the-reason-i-stayed-has-nothing-to-do-with-spe-qzdhhvk87) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"I moved my local models from Ollama to llama.cpp, and the reason I stayed has nothing to do with speed","url":"https://daily.dev/posts/i-moved-my-local-models-from-ollama-to-llama-cpp-and-the-reason-i-stayed-has-nothing-to-do-with-spe-qzdhhvk87","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/i-moved-my-local-models-from-ollama-to-llama-cpp-and-the-reason-i-stayed-has-nothing-to-do-with-spe-qzdhhvk87"},"datePublished":"2026-10-15T14:32:50.800Z","dateModified":"2026-10-15T14:41:34.302Z","description":"A long-time Ollama user explains why switching to llama.cpp improved their local AI workflow, despite similar raw token-per-second speeds. Key advantages cited...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/7d40b8b69892c308ece2fd996eab8ccd?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/7d40b8b69892c308ece2fd996eab8ccd?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"XDA Developers","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"XDA Developers","logo":"https://media.daily.dev/image/upload/s--kCg6yyAP--/f_auto,q_auto/v1774964407/logos/xda-developers?_a=BAMAMiWQ0","url":"https://daily.dev/sources/xda-developers"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/i-moved-my-local-models-from-ollama-to-llama-cpp-and-the-reason-i-stayed-has-nothing-to-do-with-spe-qzdhhvk87","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":1},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"self-hosting,local-ai,ollama,llama-cpp","timeRequired":"PT4M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"XDA Developers","item":"https://daily.dev/sources/xda-developers"},{"@type":"ListItem","position":3,"name":"I moved my local models from Ollama to llama.cpp, and the reason I stayed has nothing to do with speed"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/i-moved-my-local-models-from-ollama-to-llama-cpp-and-the-reason-i-stayed-has-nothing-to-do-with-spe-qzdhhvk87#faq","mainEntity":[{"@type":"Question","name":"What does the --n-cpu-moe flag do in llama.cpp?","acceptedAnswer":{"@type":"Answer","text":"It lets you choose how many layers of a mixture-of-experts model keep their experts in system RAM instead of GPU VRAM, giving finer control over memory placement than Ollama offers. In one test, offloading eight layers' experts to RAM increased throughput from 74.75 to 81.43 tokens per second on the same GPU and model, while Ollama has no equivalent setting. Developers tuning local model performance can track llama.cpp configuration tricks like this via daily.dev."}},{"@type":"Question","name":"Why do my local LLM benchmark numbers vary so much between runs?","acceptedAnswer":{"@type":"Answer","text":"A large swing between runs, such as a 37% difference, is often caused by the first run still loading model weights into VRAM rather than measuring steady-state performance. Llama.cpp's built-in llama-bench tool addresses this by measuring prompt processing and generation separately, repeating each test five times, and reporting the average. Anyone benchmarking local models reliably can follow discussions like this on daily.dev."}},{"@type":"Question","name":"Does Ollama support offloading specific mixture-of-experts layers to RAM like llama.cpp does?","acceptedAnswer":{"@type":"Answer","text":"No, Ollama currently has no equivalent to llama.cpp's per-layer expert offloading feature. A feature request for this capability has been open on Ollama's GitHub issue tracker since August 2025 without being implemented. Developers comparing Ollama and llama.cpp feature gaps can follow updates on daily.dev."}}]} ```