McGill-NLP/AfriqueQwen3.5-4B-Instruct-v2

VISIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:4.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Aug 27, 2026License:cc-by-4.0Architecture:Transformer Open Weights Featherless Exclusive Cold

McGill-NLP/AfriqueQwen3.5-4B-Instruct-v2 is a 4.5 billion parameter instruction-tuned language model developed by McGill-NLP, based on the Qwen3.5 architecture. This model is specifically post-trained and optimized for performance across a range of African language benchmarks, including mathematical reasoning, multilingual understanding, and translation. It demonstrates strong capabilities in tasks like AfriMGSM, AfriMMLU, and Belebele, making it suitable for applications requiring robust performance in diverse linguistic contexts.

Loading preview...

AfriqueQwen3.5-4B-Instruct-v2 Overview

AfriqueQwen3.5-4B-Instruct-v2 is an instruction-tuned language model from McGill-NLP, building upon the Qwen3.5-4B architecture. This version (v2) is a post-trained iteration of its predecessor, AfriqueQwen3.5-4B, with a focus on enhancing performance across various benchmarks relevant to African languages and contexts. The model has 4.5 billion parameters and supports a context length of 32768 tokens.

Key Capabilities and Performance

This model excels in several key areas, as demonstrated by its benchmark scores:

  • African Language Understanding: Achieves strong results on benchmarks like AfriMGSM (50.8), AfriMMLU (46.4), AfriXNLI (63.6), and Belebele (66.2).
  • Multilingual Reasoning: Shows robust performance in mathematical word problems and general language understanding across different languages, including Swahili examples provided in the usage section.
  • Competitive Performance: Outperforms several models in its size class, including Gemma-3-4B-it and tiny-aya-earth, and shows improvements over its v1 predecessor across most metrics.

Usage and Recommendations

The model is designed for deployment with vLLM and requires the --language-model-only flag due to its architecture. Specific prompting strategies and inference settings are recommended for different tasks:

  • Mathematics: Use enable_thinking: true, temperature: 0.7, top_p: 0.8, and max_tokens: 10240 with a step-by-step reasoning prompt format.
  • Other Tasks (e.g., Translation, Multiple Choice): Recommend temperature: 0 (greedy decoding) and max_tokens: 1024.

Finetunes
1 models