🎤 Micdrop

Mistral 7B

Tested as the agent, because French projects often ask for a Mistral model. Few Mistral models fit a personal computer. Mistral Small 4 is a mixture of experts with 119 billion parameters, Mistral Medium and Large are larger, and the only open weights that fit are Mistral 7B, last updated in 2024.

Like the other agents in this section, it was compared with the recommended model, Qwen3 4B Instruct, using ollama-agent.ts: the same four turns, the same prompt with the automatic prompts on, the same machine, and each model loaded before its run. The runs were done in French and in English.

TurnMistral 7B v0.3Qwen3 4B, recommended
A greetingFirst token after 0.1 to 0.3 s, 3 to 7 s of textFirst token after ~145 ms, a short answer
Asking the time, with a get_time toolNo call, an invented time, ~3 s to the first tokenTool called, the real time, 1.1 to 1.3 s
A meaningless "euh" or "uh"Answered every timeSkipped 3 times out of 4
"Merci, au revoir !"Thanked the user for “the instructions”Said goodbye without ending the call

Ollama lists tools among the capabilities of this model, but it never produced a valid tool call with the long system prompt Micdrop builds. It made up the time instead, 13 h 15 when it was almost midnight. It also treats the automatic prompts as user messages, so it answers a goodbye by thanking the user for “the detailed instructions”.

After the greeting, every turn takes about 3 seconds to start, which is already the time budget of a whole spoken turn. Use Qwen3 4B Instruct, or the Mistral API if you need a Mistral model.