Vikhrmodels/QVikhr-3-1.7B-Instruction-noreasoning
Hugging Face
TEXT GENERATIONConcurrency Cost:1Model Size:2BQuant:BF16Ctx Length:32kPublished:May 29, 2025License:apache-2.0Architecture:Transformer0.0K Open Weights Warm

QVikhr-3-1.7B-Instruction-noreasoning is a 2 billion parameter instruction-tuned language model developed by Vikhrmodels, based on Qwen3-1.7B. Fine-tuned on the GrandMaster2 dataset, it specializes in high-efficiency text processing and instruction generation in both Russian and English, with a context length of 40960 tokens. This model is optimized for precise responses and fast task execution, particularly excelling in Russian language tasks.

Loading preview...

QVikhr-3-1.7B-Instruction-noreasoning Overview

QVikhr-3-1.7B-Instruction-noreasoning is a 2 billion parameter instruction-tuned language model developed by Vikhrmodels. It is built upon the Qwen3-1.7B base model and has been specifically fine-tuned using the GrandMaster2 dataset, focusing on supervised fine-tuning (SFT) with a full fine-tune (FFT) method.

Key Capabilities

  • Bilingual Support: Designed for high-efficiency text processing and instruction generation in both Russian (RU) and English (EN).
  • Instruction Following: Optimized for instructional tasks, contextual responses, and text analysis.
  • Performance: Achieves a score of 59.2 on the Ru Arena General benchmark, outperforming its base model Qwen3-1.7B (49.7) and a 'noresoning-Qwen3-1.7B' variant (51.9).
  • Quantized Variants: Available in GGUF, MLX 4-bit, and MLX 8-bit quantized formats for broader deployment.

Good For

  • Applications requiring robust instruction generation and text analysis in Russian.
  • Integrating into user-facing applications and services that need precise and fast responses in bilingual (RU/EN) contexts.
  • Professional environments where efficient processing of textual data is critical.