Uploaded October 2026 | Updated November 2026, 3 days ago
Ollama runs llama.cpp under the hood. Same engine, same model file, but Ollama gave me 2× the speed of stock llama.cpp.
I ran Qwen 3.8 27B on my RTX 5090 locally: stock llama.cpp gave me 70 tokens per second, Ollama gave me 130.
So I checked what Ollama sends to llama.cpp. It quietly turns on the model's draft head (MTP): a small helper guesses the
next few words, and the big model checks them in one go. That explains two thirds of the gap. The rest is a different
prompt template. With the same settings, llama.cpp was about 11% faster.
The catch: Ollama throws away Qwen's own chat template, so you can't set reasoning effort, which cut my wait for the answer by a third.
My verdict: start with Ollama (fast out of the box), move to llama.cpp when you want full control.
Are you an Ollama or llama.cpp? Tell me in the comments.
Full benchmark: kgptalkie.com/tutorials/llm-benchmarking/qwen-3-8-27b-llama-cpp-speed-settings
#ollama #llamacpp #localai #qwen #rtx5090 #llm #shorts
Ollama runs llama.cpp under the hood. Same engine, same model file, but Ollama gave me 2× the speed of stock llama.cpp.
I ran Qwen 3.8 27B on my RTX 5090 locally: stock llama.cpp gave me 70 tokens per second, Ollama gave me 130.
So I checked what Ollama sends to llama.cpp. It quietly turns on the model's draft head (MTP): a small helper guesses the
next few words, and the big model checks them in one go. That explains two thirds of the gap. The rest is a different
prompt template. With the same settings, llama.cpp was about 11% faster.
The catch: Ollama throws away Qwen's own chat template, so you can't set reasoning effort, which cut my wait for the answer by a third.
My verdict: start with Ollama (fast out of the box), move to llama.cpp when you want full control.
Are you an Ollama or llama.cpp? Tell me in the comments.
Full benchmark: kgptalkie.com/tutorials/llm-benchmarking/qwen-3-8-27b-llama-cpp-speed-settings
#ollama #llamacpp #localai #qwen #rtx5090 #llm #shorts










