Explorations

Each page describes one model tested in a Micdrop call, as the agent, the transcription or the voice. A voice assistant needs things that general benchmarks do not measure: a short wait before the first word, speech generated faster than it plays, and tool calls in the right format during a conversation.

The tests run with the demo in examples/advanced. Its server/src/tests folder has one script per model, so each result can be reproduced on another machine or with a newer release.

Model	Tested as	Why it was left out
MiniCPM5-2B	The agent	Twice as fast as Qwen3 4B, but English only in practice
Mistral 7B	The agent	Slower than Qwen3 4B, and its tool calls fail
Voxtral Mini 3B	Speech to text	More accurate than Whisper base, 35 times slower
Voxtral TTS 4B	The voice	Generates slower than real time
AuK and AuK-Flash	The voice	Nine times slower than real time, even with a memory fix

All measurements were made on the machine described on the latency and memory page. The results apply to that machine, not to the models in general. A server with a large GPU, or a faster runtime for the same weights, can make any of them usable.

MiniCPM5-2B is the exception. It is not in the recommended stack, but it is recommended for an English call on a machine with little memory, as the local LLM page explains.