jsbeaudry/makandal-multiple-v2

TEXT GENERATIONPricing:Input $0.04 / Cached $0.008 / Output $0.08Concurrent Unit Cost:1Model Size:1BQuant:BF16Context Size:32kPublished:Oct 5, 2026License:gemmaArchitecture:Transformer Featherless Exclusive Cold

jsbeaudry/makandal-multiple-v2 is a 1 billion parameter causal language model, fine-tuned from google/gemma-3-1b-it, with a 32768 token context length. Developed by jsbeaudry, it is specifically optimized as a small, local brain for the Klara Kreyòl voice assistant, excelling in multi-turn conversations and tool-calling across Haitian Creole, French, Spanish, and English. Its primary differentiator is its robust tool-use capabilities and multilingual support for voice assistant interactions, delivering concise, factual responses without markdown.

Loading preview...

Overview

jsbeaudry/makandal-multiple-v2 is a 1 billion parameter language model, fine-tuned from google/gemma-3-1b-it, designed to act as the local brain for the Klara Kreyòl voice assistant. It supports multi-turn conversations and tool-calling, responding in Haitian Creole, French, Spanish, and English with concise, voice-assistant-style answers (one to three short sentences, answer first, no markdown, numbers as digits).

Key Capabilities

  • Multilingual Interaction: Fluent in Haitian Creole, French, Spanish, and English.
  • Advanced Tool Use: Integrates with eight tools including time, weather, encyclopedia, web search, calculator, personal memory, and user-saved knowledge. Achieves 93% right tool calls and 98% tool calls when needed.
  • Improved Conversation: Addresses issues from its predecessor (makandal-multiple v1) by supporting multi-turn conversations and eliminating repetitive replies (0% repeats in tests).
  • Optimized for Voice Assistants: Provides direct, factual answers suitable for voice interfaces.

Performance Highlights

  • Language Understanding: Shows improved scores on Belebele tests across all four supported languages compared to the base model and v1.
  • Tool Calling Accuracy: Demonstrates high accuracy in identifying and calling the correct tools (93%) and ensuring arguments are valid JSON (100%).
  • Mathematical Reasoning: Effectively uses the calculate tool for word problems (98-100% usage).

Good For

  • Local Voice Assistants: Ideal for small, embedded applications requiring multilingual support and tool integration.
  • Multilingual Chatbots: Suitable for conversational agents needing to interact in Haitian Creole, French, Spanish, or English.
  • Tool-Augmented LLM Applications: Projects that benefit from an LLM capable of reliably calling external functions to retrieve information or perform actions.

Finetunes
2 models