jsbeaudry/makandal-multiple-v2-GGUF

makandal-multiple-v2 — GGUF

Try it in your browser: The Trio Serie Space runs Makandal with the other two models of the series: the Klara tab is a conversation with its eight tools, by voice or text, and another tab compares it with gemma-3-1b-it (the Space runs the full-precision model). More on the Makandal page of thetrio.space.

jsbeaudry/makandal-multiple-v2 for llama.cpp: Gemma 3 1B taught to answer in Haitian Creole, French, Spanish and English the way a voice assistant speaks, to hold a conversation, and to call tools (time, weather, encyclopedia, search, calculator, memory, saved knowledge). It is the local brain of Klara.

quant	size	speed on an M3 Pro (Metal)
Q8_0	1.07 GB	~91 tokens/s
Q4_K_M	0.81 GB	~101 tokens/s

Q8_0 was converted from the trained weights on the training machine; Q4_K_M was quantized from the 16-bit weights. Both were checked in llama-server before publishing: the same tool calls on weather, encyclopedia, percentage and small-talk requests, and the same kind of spoken answer. Use Q8_0 unless size is binding: at 1B, 4-bit quantization costs facts first.

See the main card for the results and the known limits: tool calls are reliable on a first request and much less so later in a conversation, a gap a v2.1 is planned to close.

Use

--jinja is needed for tools: it runs the chat template stored in the file and turns the <tool_call>{…}</tool_call> blocks the model writes into OpenAI tool_calls.

llama-server -m makandal-multiple-v2-Q8_0.gguf -c 8192 --jinja

curl http://127.0.0.1:8080/v1/chat/completions -H 'Content-Type: application/json' -d '{
  "messages": [{"role": "system", "content": "Ou rele Klara. Ou se yon asistan vwa ki pale kreyòl ayisyen.\n"},
               {"role": "user", "content": "Ki tan l ap fè Okap jodi a?"}],
  "tools": [{"type": "function", "function": {"name": "weather", "description": "The weather now in a place.",
             "parameters": {"type": "object", "properties": {"place": {"type": "string"}}, "required": ["place"]}}}]}'

After a tool result, send the call back with a short line before it in the assistant message (for example "content": "Kite m gade sa."): the model was trained that way, and with an empty one it tends to call another tool instead of answering.