← All video evidence
VIDEO DOCUMENTEDSOURCE CHECKED

Why can llama.cpp server outperform a simple Ollama test?

Alex Ziskind compares local serving approaches and shows why a single chat response is not enough to judge an agent server. The test examines Ollama, llama.cpp server, remote access, parallel requests, and aggregate throughput.

Video by Alex Ziskind · Reviewed August 22, 2026

The video remains on YouTube. It belongs to its creator, and LocalLens does not host a copy.

LOCALLENS REVIEW

What this evidence supports

The video correctly highlights the difference between single-request speed and total server throughput. Its public benchmark code makes the method inspectable. The exact speed figures remain specific to Alex's model, Mac Studio, server flags, and concurrency settings, so LocalLens presents them as reported results rather than universal engine rankings.

Medium to high confidenceThis is not a LocalLens reproduction.
HARDWARE SHOWN OR REPORTED

Computer setup

  • Mac Studio used as the local model server
  • Laptop used as a remote client
  • Exact Mac Studio chip and memory need confirmation
SOFTWARE SHOWN OR REPORTED

Model and engine

  • Ollama
  • llama.cpp server
  • 34B-class GGUF model
  • Public llama-throughput-lab benchmark harness
MEASUREMENTS

Results recorded in the video

Ollama exampleAbout 100 tokens per secondReported in video
llama.cpp server, single requestAbout 124 tokens per secondReported in video
Remote request to Mac StudioAbout 120 tokens per secondReported in video
Concurrent servingHigher aggregate throughput demonstratedReported in video
CROSS-CHECKED

What the evidence supports

  • The creator published the llama-throughput-lab code used to run single, concurrent, round-robin, and sweep tests.
  • The harness exposes parallelism, context, threads, batch settings, model path, and request concurrency.
  • llama.cpp supports separate prompt-processing, text-generation, and combined benchmark measurements.
STILL UNCONFIRMED

What we cannot claim

  • LocalLens did not reproduce the reported speeds.
  • The exact Mac Studio chip, unified memory, model file, quantization, context, Ollama version, and llama.cpp commit need confirmation.
  • Aggregate throughput must not be presented as the speed one user sees in a single response.
SOURCE RECORD

Sources used for this review