Vims2-7B Overview
Vims2-7B is a 7.6 billion parameter large language model developed by specialv, built upon the Qwen 2.5 architecture. It leverages the Task Arithmetic merging method to combine the strengths of Qwen 2.5 Instruct and Coder variants, creating a specialized model that excels in logical reasoning, mathematical problem-solving, and coding. This model is proficient in both Italian and English, offering superior instruction-following capabilities in both languages.
Key Capabilities
- Enhanced Logical Reasoning & Math: Achieves a simulated 100.0% on GSM8K (Math) benchmarks, indicating strong performance in quantitative tasks.
- Proficient Coding: Optimized for algorithmic logic, making it suitable for code-related applications.
- Dual-Language Instruction Following: Excels in understanding and executing instructions in both Italian and English.
- Efficient Inference: Utilizes Grouped Query Attention (GQA) to reduce KV Cache memory usage, enabling faster inference and larger batch sizes on consumer GPUs.
- Long Context Window: Supports a substantial context window of 131,072 tokens (128k).
Good For
- Applications requiring strong mathematical and logical reasoning.
- Code generation and understanding tasks.
- Multilingual chatbots or assistants needing high instruction adherence in Italian and English.
- Deployment on consumer-grade GPUs, optimized for 4-bit quantization to fit within 16GB VRAM.